# LLM Tree — Models # # The single source of truth for the visualization at /atlas/llm-tree. # Schema: see schema.md. Branch placement: see placement-rules.md. # # To add a new model, see README.md ("How to add a new model in 15 minutes"). # The validator (`npm run validate:llm-tree`) blocks the build on any # integrity error or redaction-blocklist hit. # # Last updated: 2026-04-25 models: # ============================================================ # FOUNDATIONAL (pre-transformer language representations) # ============================================================ - id: word2vec name: Word2Vec vendor: Google release_date: "2013-01-16" year_bucket: 2013 branch: foundational secondary_tags: [open_source, foundational, embedding] parents: [] influences: [] openness: open_source parameters: "n/a" context: "n/a" modalities: [text] architecture_notes: > CBOW and skip-gram word embedding models trained via hierarchical softmax / negative sampling on shallow neural network. capabilities: [] placement_confidence: high notable: > Established distributed word representations as the foundation of modern NLP. Trained 300-dim vectors over 1.6B words in under 1 day on a single CPU. source_refs: [src-word2vec-paper] - id: glove name: GloVe vendor: Stanford NLP release_date: "2014-10-25" year_bucket: 2014 branch: foundational secondary_tags: [open_source, foundational, embedding] parents: [word2vec] influences: [word2vec] openness: open_source parameters: "n/a" context: "n/a" modalities: [text] architecture_notes: > Global log-bilinear regression over word-word co-occurrence matrix. Combines global matrix factorization with local context-window training. capabilities: [] placement_confidence: high notable: > 75% accuracy on word analogy tasks at release; ~30K citations. Pennington/Socher/Manning, EMNLP 2014. source_refs: [src-glove-paper] - id: fasttext name: FastText vendor: Meta release_date: "2016-07-15" year_bucket: 2016 branch: foundational secondary_tags: [open_source, foundational, embedding, multilingual] parents: [word2vec] influences: [word2vec] openness: open_source parameters: "n/a" context: "n/a" modalities: [text] architecture_notes: > Skip-gram extension where each word is the sum of its character n-gram vectors. Handles out-of-vocabulary words and morphologically rich languages. capabilities: [] placement_confidence: high notable: > Mikolov's second-act embedding work at Facebook. Introduced subword information; foundation of Facebook's open-source fastText library. source_refs: [src-fasttext-paper] - id: elmo name: ELMo vendor: Allen Institute for AI release_date: "2018-02-15" year_bucket: 2018 branch: foundational secondary_tags: [open_source, foundational] parents: [word2vec, glove] influences: [] openness: open_source parameters: "~94M (biLM)" context: "n/a" modalities: [text] architecture_notes: > Bidirectional LSTM language model; word vectors are learned linear combinations of internal biLM states, capturing context. capabilities: [] placement_confidence: high notable: > First widely-adopted contextualized word representation. NAACL 2018 best paper. Last major pre-transformer milestone before BERT. source_refs: [src-elmo-paper] - id: ulmfit name: ULMFiT vendor: fast.ai release_date: "2018-01-18" year_bucket: 2018 branch: foundational secondary_tags: [open_source, foundational] parents: [] influences: [] openness: open_source parameters: "~24M (AWD-LSTM)" context: "n/a" modalities: [text] architecture_notes: > 3-layer AWD-LSTM language model with discriminative fine-tuning, slanted triangular learning rates, and gradual unfreezing for downstream tasks. capabilities: [] placement_confidence: high notable: > Established the pretrain-then-finetune paradigm for NLP that GPT and BERT generalized later in 2018. Howard and Ruder, ACL 2018. source_refs: [src-ulmfit-paper] # ============================================================ # ENCODER-ONLY # ============================================================ - id: bert-base name: BERT (base) vendor: Google release_date: "2018-10-11" year_bucket: 2018 branch: encoder_only secondary_tags: [open_source, foundational] parents: [] influences: [] tree_parent: transformer openness: open_source parameters: "110M" context: "512" modalities: [text] architecture_notes: > Bidirectional encoder, masked language modeling (MLM) plus next sentence prediction (NSP) pretraining. capabilities: [] placement_confidence: high notable: > Defined the encoder-only branch; first widely-deployed bidirectional MLM pretraining objective. source_refs: [src-bert-paper] - id: bert-large name: BERT (large) vendor: Google release_date: "2018-10-11" year_bucket: 2018 branch: encoder_only secondary_tags: [open_source, foundational] parents: [bert-base] influences: [] openness: open_source parameters: "340M" context: "512" modalities: [text] architecture_notes: > 24-layer, 1024-hidden, 16 attention heads variant of BERT. capabilities: [] placement_confidence: high notable: > Topped GLUE and SQuAD on release. source_refs: [src-bert-paper] - id: roberta name: RoBERTa vendor: Meta release_date: "2019-07-26" year_bucket: 2019 branch: encoder_only secondary_tags: [open_source] parents: [bert-base] influences: [] openness: open_source parameters: "125M (base) / 355M (large)" context: "512" modalities: [text] architecture_notes: > Same architecture as BERT; trained 10x longer on 10x more data, dropped NSP, dynamic masking. capabilities: [] placement_confidence: high notable: > Showed BERT was significantly under-trained. source_refs: [src-roberta-paper] - id: albert name: ALBERT vendor: Google release_date: "2019-09-26" year_bucket: 2019 branch: encoder_only secondary_tags: [open_source, small] parents: [bert-base] influences: [] openness: open_source parameters: "12M / 18M / 60M / 235M (xxlarge)" context: "512" modalities: [text] architecture_notes: > Cross-layer parameter sharing plus factorized embedding for memory efficiency. capabilities: [] placement_confidence: high notable: > Lite BERT; xxlarge has fewer parameters than BERT-large but outperforms it on GLUE/SQuAD/RACE. source_refs: [src-albert-paper] - id: distilbert name: DistilBERT vendor: HuggingFace release_date: "2019-10-02" year_bucket: 2019 branch: encoder_only secondary_tags: [open_source, distilled, small, edge] parents: [bert-base] influences: [] openness: open_source parameters: "66M" context: "512" modalities: [text] architecture_notes: > Knowledge distillation from BERT-base teacher; triple loss (LM, distillation, cosine). capabilities: [] placement_confidence: high notable: > 40% smaller, 60% faster, retains ~97% of BERT's GLUE performance. Foundational for on-device NLP. source_refs: [src-distilbert-paper] - id: xlnet name: XLNet vendor: CMU + Google release_date: "2019-06-19" year_bucket: 2019 branch: encoder_only secondary_tags: [open_source] parents: [bert-base] influences: [] openness: open_source parameters: "110M (base) / 340M (large)" context: "512" modalities: [text] architecture_notes: > Permutation language modeling combines AR factorization with bidirectional context; built on Transformer-XL. capabilities: [] placement_confidence: high notable: > First major bridge between AR (GPT-style) and bidirectional (BERT-style) pretraining. source_refs: [src-xlnet-paper] - id: electra name: ELECTRA vendor: Google release_date: "2020-03-23" year_bucket: 2020 branch: encoder_only secondary_tags: [open_source, small] parents: [bert-base] influences: [] openness: open_source parameters: "14M (small) / 110M (base) / 335M (large)" context: "512" modalities: [text] architecture_notes: > Replaced token detection (RTD): discriminator distinguishes real vs. generator-replaced tokens; uses every token, unlike MLM. capabilities: [] placement_confidence: high notable: > Matches RoBERTa/XLNet quality at a fraction of compute; ELECTRA-small runs on a single GPU. source_refs: [src-electra-paper] - id: ernie-3 name: ERNIE 3.0 vendor: Baidu release_date: "2021-07-05" year_bucket: 2021 branch: encoder_only secondary_tags: [open_source, multilingual, china] parents: [] influences: [bert-base, t5] openness: open_source parameters: "10B (ERNIE 3.0); 260B (ERNIE 3.0 Titan)" context: "512" modalities: [text] architecture_notes: > Unifies auto-regressive and auto-encoding; trained on 4TB of plain text plus knowledge graph. capabilities: [] placement_confidence: medium notable: > Topped SuperGLUE (90.6%) — first to exceed human baseline on July 3, 2021. Titan was the largest Chinese dense LM at the time. source_refs: [src-ernie-3-paper] - id: text-embedding-3 name: text-embedding-3 vendor: OpenAI release_date: "2024-01-25" year_bucket: 2024 branch: encoder_only secondary_tags: [closed, embedding] parents: [] influences: [gpt-3] openness: closed parameters: "undisclosed" context: "8192" modalities: [text] architecture_notes: > Embedding-only transformer family. text-embedding-3-large supersedes text-embedding-ada-002. Variable output dimensions via Matryoshka representation. capabilities: [] placement_confidence: medium notable: > OpenAI's only major encoder-only entry; included for architectural completeness. source_refs: [src-text-embedding-3-blog] # ============================================================ # ENCODER-DECODER # ============================================================ - id: t5 name: T5 vendor: Google release_date: "2019-10-23" year_bucket: 2019 branch: encoder_decoder secondary_tags: [open_source, foundational] parents: [] influences: [] tree_parent: transformer openness: open_source parameters: "60M / 220M / 770M / 3B / 11B" context: "512" modalities: [text] architecture_notes: > Unified text-to-text framework; span-corruption denoising pretraining on C4 (Colossal Clean Crawled Corpus). capabilities: [] placement_confidence: high notable: > Defined the modern encoder-decoder line; every NLP task framed as text-to-text. Anchored Switch Transformer, UL2, FLAN-T5. source_refs: [src-t5-paper] - id: bart name: BART vendor: Meta release_date: "2019-10-29" year_bucket: 2019 branch: encoder_decoder secondary_tags: [open_source] parents: [] influences: [bert-base] openness: open_source parameters: "140M (base) / 400M (large)" context: "1024" modalities: [text] architecture_notes: > Bidirectional encoder plus autoregressive decoder; pretrained on 5 different noising functions including text infilling and sentence permutation. capabilities: [] placement_confidence: high notable: > Combines BERT (bidirectional encoder) and GPT (autoregressive decoder). Workhorse for summarization. source_refs: [src-bart-paper] - id: t0 name: T0 vendor: BigScience release_date: "2021-10-15" year_bucket: 2021 branch: encoder_decoder secondary_tags: [open_source, multilingual] parents: [t5] influences: [] openness: open_source parameters: "3B / 11B" context: "1024" modalities: [text] architecture_notes: > T5-LM-adapted base, then fine-tuned on a multitask mixture of NLP datasets converted to natural-language prompts. capabilities: [] placement_confidence: high notable: > Outperformed GPT-3 (175B) zero-shot on many tasks while being 16x smaller. source_refs: [src-t0-paper] - id: ul2 name: UL2 vendor: Google release_date: "2022-05-10" year_bucket: 2022 branch: encoder_decoder secondary_tags: [open_source] parents: [t5] influences: [] openness: open_source parameters: "20B" context: "2048" modalities: [text] architecture_notes: > Mixture-of-Denoisers (MoD) pretraining unifying causal LM, prefix LM, and span corruption; mode-switching tags at inference. capabilities: [] placement_confidence: high notable: > Beat GPT-3 zero-shot SuperGLUE; later instruction-tuned as Flan-UL2. source_refs: [src-ul2-paper] - id: glm name: GLM (original) vendor: Tsinghua THUDM release_date: "2021-03-18" year_bucket: 2021 branch: encoder_decoder secondary_tags: [open_source, bilingual, china] parents: [] influences: [] openness: open_source parameters: "110M to 130B (GLM-130B)" context: "2K" modalities: [text] architecture_notes: > Autoregressive blank-filling objective unifies NLU plus conditional NLG plus unconditional NLG in one framework. capabilities: [] placement_confidence: high notable: > Bilingual (English plus Chinese); GLM-130B (October 2022) outperformed GPT-3 175B and BLOOM-176B on LAMBADA/MMLU. Anchored ChatGLM/GLM-4. source_refs: [src-glm-paper] - id: flan-t5 name: FLAN-T5 vendor: Google release_date: "2022-10-20" year_bucket: 2022 branch: encoder_decoder secondary_tags: [open_source] parents: [t5] influences: [t0] openness: open_source parameters: "80M / 250M / 780M / 3B / 11B" context: "1024" modalities: [text] architecture_notes: > T5 instruction-finetuned on 1,836 tasks (Flan Collection) with chain-of-thought data. capabilities: [] placement_confidence: high notable: > Reference instruction-tuned encoder-decoder. Showed instruction tuning plus CoT scaling generalizes. source_refs: [src-flan-t5-paper] # ============================================================ # DECODER-ONLY # ============================================================ # ---- OpenAI GPT lineage ---- - id: gpt-1 name: GPT-1 vendor: OpenAI release_date: "2018-06-11" year_bucket: 2018 branch: decoder_only secondary_tags: [closed, foundational] parents: [] influences: [] tree_parent: transformer openness: closed parameters: "117M" context: "512" modalities: [text] architecture_notes: > First decoder-only transformer trained generatively on BooksCorpus. Established the unsupervised pre-training plus supervised fine-tuning recipe. capabilities: [] placement_confidence: high notable: > Origin of the GPT family. Shipped as the "Improving Language Understanding" research paper, not a product. source_refs: [src-gpt-1-paper] - id: gpt-2 name: GPT-2 vendor: OpenAI release_date: "2019-02-14" year_bucket: 2019 branch: decoder_only secondary_tags: [open_weights, historical_frontier] parents: [gpt-1] influences: [] openness: open_weights parameters: "1.5B" context: "1024" modalities: [text] architecture_notes: > 13x scale-up over GPT-1; demonstrated zero-shot task transfer at scale. Initially withheld for safety; full 1.5B weights released Nov 5, 2019. capabilities: [open_weights] placement_confidence: high notable: > First major "too dangerous to release" moment in modern LLMs. The staged release set the template for capability/safety tradeoffs. source_refs: [src-gpt-2-blog] - id: gpt-3 name: GPT-3 vendor: OpenAI release_date: "2020-06-11" year_bucket: 2020 branch: decoder_only secondary_tags: [closed, historical_frontier] parents: [gpt-2] influences: [] openness: closed parameters: "175B" context: "2048" modalities: [text] architecture_notes: > ~116x scale-up over GPT-2. Demonstrated few-shot in-context learning as an emergent capability of scale. Davinci variant became the workhorse. capabilities: [] placement_confidence: high notable: > The scaling-laws moment. First closed-API LLM at this size; launched the OpenAI API. source_refs: [src-gpt-3-paper] - id: gpt-3-5-turbo name: GPT-3.5 Turbo vendor: OpenAI release_date: "2022-11-30" year_bucket: 2022 branch: decoder_only secondary_tags: [closed, historical_frontier, tool_using] parents: [gpt-3] influences: [] openness: closed parameters: "~20B (estimated; undisclosed)" context: "4K (later 16K)" modalities: [text] architecture_notes: > RLHF/PPO applied to a smaller GPT-3-class model. The "ChatGPT" release ran on this; function calling added in mid-2023. capabilities: [tools] placement_confidence: high notable: > Launched ChatGPT (Nov 30, 2022), the fastest consumer software to 100M users. Made RLHF chat tuning the industry default. source_refs: [src-chatgpt-blog] - id: gpt-4 name: GPT-4 vendor: OpenAI release_date: "2023-03-14" year_bucket: 2023 branch: decoder_only secondary_tags: [closed, historical_frontier, tool_using] parents: [gpt-3-5-turbo] influences: [] openness: closed parameters: "undisclosed" context: "8K (later 32K variants)" modalities: [text] architecture_notes: > Major capability jump. OpenAI publicly disclosed almost nothing about architecture. Vision input rolled out separately as GPT-4V (Sept 2023). capabilities: [tools] placement_confidence: high notable: > Set the bar for closed-frontier models for ~14 months. Retired from ChatGPT April 30, 2025. source_refs: [src-gpt-4-blog] - id: gpt-4-turbo name: GPT-4 Turbo vendor: OpenAI release_date: "2023-11-06" year_bucket: 2023 branch: decoder_only secondary_tags: [closed, long_context, tool_using] parents: [gpt-4] influences: [] openness: closed parameters: "undisclosed" context: "128K" modalities: [text, image] architecture_notes: > Cheaper, faster, 128K-context iteration on GPT-4. Vision rolled into the base model. capabilities: [tools, multimodal, long_context] placement_confidence: high notable: > Established 128K as the de-facto context-window expectation for frontier models. source_refs: [src-gpt-4-turbo-blog] - id: gpt-4-5 name: GPT-4.5 vendor: OpenAI release_date: "2025-02-27" year_bucket: 2025 branch: decoder_only secondary_tags: [closed, historical_frontier, tool_using] parents: [gpt-4-turbo] influences: [gpt-4o] openness: closed parameters: "undisclosed" context: "128K" modalities: [text, image] architecture_notes: > Pretraining-scaling research preview; explicitly NOT a reasoning model. Last major pre-train-only flagship before GPT-5 unified. capabilities: [tools, multimodal, long_context] placement_confidence: high notable: > Marked the end of pure-pretraining-scaling as OpenAI's frontier strategy. The next flagship (GPT-5) was unified with reasoning. source_refs: [src-gpt-4-5-blog] # ---- Anthropic Claude lineage (decoder_only era) ---- - id: claude-1 name: Claude 1 vendor: Anthropic release_date: "2023-03-14" year_bucket: 2023 branch: decoder_only secondary_tags: [closed, historical_frontier, long_context] parents: [] influences: [] openness: closed parameters: "undisclosed" context: "9K, expanded to 100K (May 2023)" modalities: [text] architecture_notes: > Dense decoder; trained with Constitutional AI / RLHF; no public weights. capabilities: [long_context] placement_confidence: high notable: > Anthropic's first public model. May 2023 100K context window was an industry first. source_refs: [src-claude-1-blog] - id: claude-instant-1 name: Claude Instant 1 vendor: Anthropic release_date: "2023-03-14" year_bucket: 2023 branch: decoder_only secondary_tags: [closed, small, long_context] parents: [claude-1] influences: [] openness: closed parameters: "undisclosed (smaller than Claude 1)" context: "100K" modalities: [text] architecture_notes: > Distilled / smaller-class dense decoder; faster and cheaper sibling of Claude 1. capabilities: [long_context] placement_confidence: high notable: > Anthropic's first cost/latency tier; direct ancestor of the Haiku branch. source_refs: [src-claude-instant-blog] - id: claude-2 name: Claude 2 vendor: Anthropic release_date: "2023-07-11" year_bucket: 2023 branch: decoder_only secondary_tags: [closed, long_context] parents: [claude-1] influences: [] openness: closed parameters: "undisclosed" context: "100K" modalities: [text] architecture_notes: > Dense decoder; first public API access; longer outputs than Claude 1. capabilities: [long_context] placement_confidence: high notable: > First Claude broadly available via API; coding plus math gains over Claude 1. source_refs: [src-claude-2-blog] - id: claude-2-1 name: Claude 2.1 vendor: Anthropic release_date: "2023-11-21" year_bucket: 2023 branch: decoder_only secondary_tags: [closed, long_context, tool_using] parents: [claude-2] influences: [] openness: closed parameters: "undisclosed" context: "200K" modalities: [text] architecture_notes: > Dense decoder; introduced 200K context, system prompts, and tool use beta. capabilities: [tools, long_context] placement_confidence: high notable: > First Claude with 200K context; first tool-use beta. Retired July 21, 2025. source_refs: [src-claude-2-1-blog] - id: claude-3-haiku name: Claude 3 Haiku vendor: Anthropic release_date: "2024-03-13" year_bucket: 2024 branch: decoder_only secondary_tags: [closed, small, long_context, multimodal] parents: [claude-2-1] influences: [claude-instant-1] openness: closed parameters: "undisclosed" context: "200K" modalities: [text, image] architecture_notes: > Dense decoder; vision (image input) added at the family level; cheapest/fastest tier. capabilities: [tools, multimodal, long_context] placement_confidence: high notable: > First multimodal Claude tier. source_refs: [src-claude-3-blog] - id: claude-3-sonnet name: Claude 3 Sonnet vendor: Anthropic release_date: "2024-03-04" year_bucket: 2024 branch: decoder_only secondary_tags: [closed, long_context, multimodal] parents: [claude-2-1] influences: [] openness: closed parameters: "undisclosed" context: "200K" modalities: [text, image] architecture_notes: > Dense decoder; balanced-tier of the Claude 3 family; vision-capable. capabilities: [tools, multimodal, long_context] placement_confidence: high notable: > Balanced cost/intelligence tier of Claude 3. source_refs: [src-claude-3-blog] - id: claude-3-opus name: Claude 3 Opus vendor: Anthropic release_date: "2024-03-04" year_bucket: 2024 branch: decoder_only secondary_tags: [closed, historical_frontier, long_context, multimodal] parents: [claude-2-1] influences: [] openness: closed parameters: "undisclosed" context: "200K" modalities: [text, image] architecture_notes: > Dense decoder; flagship tier of Claude 3; vision-capable. capabilities: [tools, multimodal, long_context] placement_confidence: high notable: > Anthropic's first model to surpass GPT-4 on multiple benchmarks. source_refs: [src-claude-3-blog] - id: claude-3-5-sonnet name: Claude 3.5 Sonnet vendor: Anthropic release_date: "2024-06-20" year_bucket: 2024 branch: decoder_only secondary_tags: [closed, long_context, multimodal, agentic, computer_use] parents: [claude-3-sonnet] influences: [] openness: closed parameters: "undisclosed" context: "200K" modalities: [text, image] architecture_notes: > Dense decoder; outperformed Claude 3 Opus on most benchmarks at Sonnet pricing. Computer use API added in Oct 22, 2024 update. capabilities: [tools, multimodal, long_context, computer_use] placement_confidence: high notable: > Introduced computer-use API in October 2024 — first agentic GUI-control SKU. source_refs: [src-claude-3-5-sonnet-blog, src-claude-computer-use-blog] - id: claude-3-5-haiku name: Claude 3.5 Haiku vendor: Anthropic release_date: "2024-11-04" year_bucket: 2024 branch: decoder_only secondary_tags: [closed, small, long_context, tool_using, multimodal] parents: [claude-3-haiku] influences: [] openness: closed parameters: "undisclosed" context: "200K" modalities: [text, image] architecture_notes: > Dense decoder; text-only at launch, vision added Feb 24, 2025 alongside Claude 3.7 Sonnet. capabilities: [tools, long_context, multimodal] placement_confidence: high notable: > Matched Claude 3 Opus performance at Haiku-tier cost. source_refs: [src-claude-computer-use-blog] # ---- Google decoder-only ---- - id: lamda name: LaMDA vendor: Google release_date: "2021-05-18" year_bucket: 2021 branch: decoder_only secondary_tags: [closed] parents: [] influences: [] openness: closed parameters: "Up to 137B" context: "Not disclosed" modalities: [text] architecture_notes: > Decoder-only transformer pretrained on 1.56T words of dialog plus web; fine-tuned for safety, grounding via tool/IR calls. capabilities: [] placement_confidence: high notable: > Powered the original Bard at March 2023 launch. source_refs: [src-lamda-paper] - id: palm name: PaLM vendor: Google release_date: "2022-04-04" year_bucket: 2022 branch: decoder_only secondary_tags: [closed, historical_frontier] parents: [] influences: [] openness: closed parameters: "8B / 62B / 540B" context: "2048" modalities: [text] architecture_notes: > Dense decoder-only; trained on 6,144 TPU v4 chips via Pathways at 57.8% hardware FLOPs utilization; 780B token mix. capabilities: [] placement_confidence: high notable: > Largest dense LM at release; chain-of-thought reasoning emerged at 540B scale. source_refs: [src-palm-paper] - id: palm-2 name: PaLM 2 vendor: Google release_date: "2023-05-10" year_bucket: 2023 branch: decoder_only secondary_tags: [closed, multilingual] parents: [palm] influences: [chinchilla] openness: closed parameters: "Undisclosed; sizes Gecko / Otter / Bison / Unicorn" context: "8K" modalities: [text] architecture_notes: > Compute-optimal scaling per Chinchilla laws; multilingual training over 100+ languages; specialized variants Med-PaLM 2, Sec-PaLM. capabilities: [] placement_confidence: high notable: > Powered Bard from May 2023 to Dec 2023 transition. source_refs: [src-palm-2-blog] - id: chinchilla name: Chinchilla vendor: DeepMind release_date: "2022-03-29" year_bucket: 2022 branch: decoder_only secondary_tags: [research_only, foundational] parents: [] influences: [] openness: research_only parameters: "70B (trained on 1.3T tokens)" context: "2048" modalities: [text] architecture_notes: > 80 layers, 64 attention heads; same compute as Gopher (280B) but tokens scaled equally with parameters. capabilities: [] placement_confidence: high notable: > Established compute-optimal training law: tokens and parameters should scale equally. Outperformed Gopher (280B) and GPT-3 (175B) on MMLU. source_refs: [src-chinchilla-paper] status: active - id: sparrow name: Sparrow vendor: DeepMind release_date: "2022-09-28" year_bucket: 2022 branch: decoder_only secondary_tags: [research_only] parents: [chinchilla] influences: [] openness: research_only parameters: "70B (Chinchilla backbone, fine-tuned)" context: "2048" modalities: [text] architecture_notes: > Chinchilla 70B fine-tuned with RLHF using rule-conditional reward models plus evidence retrieval; trained against 23 explicit dialogue rules. capabilities: [] placement_confidence: high notable: > Research-only; never productized. Influenced safety methodology in Bard / Gemini and broader RLHF practice. source_refs: [src-sparrow-paper] status: research_only display_in_tree: false - id: gemma-1 name: Gemma 1 vendor: Google release_date: "2024-02-21" year_bucket: 2024 branch: decoder_only secondary_tags: [open_weights, small, edge] parents: [] influences: [gemini-1] openness: open_weights parameters: "2B / 7B" context: "8K" modalities: [text] architecture_notes: > Open-weights decoder-only built on Gemini 1.0 research and infrastructure. capabilities: [open_weights] placement_confidence: high notable: > Google's open-weights re-entry. source_refs: [src-gemma-1-blog] - id: gemma-2 name: Gemma 2 vendor: Google release_date: "2024-06-27" year_bucket: 2024 branch: decoder_only secondary_tags: [open_weights, small] parents: [gemma-1] influences: [gemini-1-5] openness: open_weights parameters: "2B / 9B / 27B" context: "8K" modalities: [text] architecture_notes: > Sliding-window attention, soft logit capping, knowledge distillation from larger teacher. capabilities: [open_weights] placement_confidence: high notable: > 27B variant matched Llama-3-70B on many benchmarks at ~1/3 the parameters. source_refs: [src-gemma-1-blog] # ---- Meta Llama lineage ---- - id: llama-1 name: LLaMA 1 vendor: Meta AI release_date: "2023-02-24" year_bucket: 2023 branch: decoder_only secondary_tags: [research_only, foundational] parents: [] influences: [chinchilla, palm] openness: research_only parameters: "7B, 13B, 33B, 65B" context: "2K" modalities: [text] architecture_notes: > RMSNorm, SwiGLU, RoPE; trained 1.0-1.4T tokens; weights leaked on 4chan March 3, 2023. capabilities: [] placement_confidence: high notable: > Set the recipe (RMSNorm/SwiGLU/RoPE) every Western open model now follows. source_refs: [src-llama-1-paper] - id: llama-2 name: Llama 2 vendor: Meta AI release_date: "2023-07-18" year_bucket: 2023 branch: decoder_only secondary_tags: [open_weights] parents: [llama-1] influences: [] openness: open_weights parameters: "7B, 13B, 70B" context: "4K" modalities: [text] architecture_notes: > GQA on 70B; trained 2T tokens; chat models RLHF'd. capabilities: [open_weights] placement_confidence: high notable: > First commercially-licensed Llama; legitimized open-weights ecosystem outside research. source_refs: [src-llama-2-news] - id: code-llama name: Code Llama vendor: Meta AI release_date: "2023-08-24" year_bucket: 2023 branch: decoder_only secondary_tags: [open_weights, code_specialist] parents: [llama-2] influences: [] openness: open_weights parameters: "7B, 13B, 34B, 70B" context: "100K (effective)" modalities: [text] architecture_notes: > Continued pretraining on 500B code tokens; FIM for 7B/13B; Python and Instruct variants. capabilities: [open_weights] placement_confidence: high notable: > Anchor of the open-coder branch; precursor to most open coding fine-tunes. source_refs: [src-code-llama-blog] - id: llama-3 name: Llama 3 vendor: Meta AI release_date: "2024-04-18" year_bucket: 2024 branch: decoder_only secondary_tags: [open_weights, historical_frontier] parents: [llama-2] influences: [] openness: open_weights parameters: "8B, 70B" context: "8K" modalities: [text] architecture_notes: > GQA across both sizes; 128K-token tokenizer; trained 15T tokens. capabilities: [open_weights] placement_confidence: high notable: > Reset the open-weight quality bar; 8B beat Llama 2 70B on many benchmarks. source_refs: [src-llama-3-blog] - id: llama-3-1 name: Llama 3.1 vendor: Meta AI release_date: "2024-07-23" year_bucket: 2024 branch: decoder_only secondary_tags: [open_weights, historical_frontier, multilingual, long_context] parents: [llama-3] influences: [] openness: open_weights parameters: "8B, 70B, 405B" context: "128K" modalities: [text] architecture_notes: > 405B is the first openly-released GPT-4-class dense model; 8 languages; permissive license for synthetic-data generation. capabilities: [open_weights, long_context] placement_confidence: high notable: > 405B closed the gap to GPT-4 in open weights. source_refs: [src-llama-3-1-blog] - id: llama-3-2-edge name: Llama 3.2 (1B / 3B edge) vendor: Meta AI release_date: "2024-09-25" year_bucket: 2024 branch: decoder_only secondary_tags: [open_weights, small, edge, long_context] parents: [llama-3-1] influences: [] openness: open_weights parameters: "1B, 3B" context: "128K" modalities: [text] architecture_notes: > Distilled/pruned from larger Llama 3.1; optimized for on-device and mobile. capabilities: [open_weights, long_context] placement_confidence: high notable: > First Meta entries in the on-device branch. source_refs: [src-llama-3-2-blog] - id: llama-3-3-70b name: Llama 3.3 70B vendor: Meta AI release_date: "2024-12-06" year_bucket: 2024 branch: decoder_only secondary_tags: [open_weights, multilingual, long_context] parents: [llama-3-1] influences: [] openness: open_weights parameters: "70B" context: "128K" modalities: [text] architecture_notes: > Post-training only update; uses online RLHF; targets 405B-class quality at 70B size. capabilities: [open_weights, long_context] placement_confidence: high notable: > Showed post-training alone could push 70B to near-405B quality. source_refs: [src-llama-3-3-card] # ---- Mistral lineage ---- - id: mistral-7b name: Mistral 7B vendor: Mistral AI release_date: "2023-09-27" year_bucket: 2023 branch: decoder_only secondary_tags: [open_weights, small, eu] parents: [llama-2] influences: [] openness: open_weights parameters: "7.3B" context: "8K (sliding window 4K)" modalities: [text] architecture_notes: > GQA plus Sliding-Window Attention; Apache 2.0; outperformed Llama 2 13B. capabilities: [open_weights] placement_confidence: high notable: > Defined the open small-model frontier; Apache 2.0 set the licensing tone for European OSS LLMs. source_refs: [src-mistral-7b-blog] - id: mistral-large-1 name: Mistral Large vendor: Mistral AI release_date: "2024-02-26" year_bucket: 2024 branch: decoder_only secondary_tags: [closed, eu] parents: [mistral-7b] influences: [] openness: closed parameters: "undisclosed (~120B est.)" context: "32K" modalities: [text] architecture_notes: > Mistral's first closed flagship; included as lineage anchor for Large 2. capabilities: [] placement_confidence: medium notable: > Marked Mistral's split into closed flagships and open mid-weights. source_refs: [src-mistral-large-blog] - id: codestral-22b name: Codestral 22B vendor: Mistral AI release_date: "2024-05-29" year_bucket: 2024 branch: decoder_only secondary_tags: [code_specialist, eu] parents: [mistral-7b] influences: [code-llama] openness: mixed parameters: "22.2B" context: "32K" modalities: [text] architecture_notes: > MNPL-0.1 (non-production for commercial use); 81.1% HumanEval. capabilities: [] placement_confidence: high notable: > Mistral's entry on the open-coder branch. source_refs: [src-codestral-blog] - id: mistral-nemo-12b name: Mistral NeMo 12B vendor: Mistral AI / NVIDIA release_date: "2024-07-18" year_bucket: 2024 branch: decoder_only secondary_tags: [open_weights, multilingual, long_context, eu] parents: [mistral-7b] influences: [] openness: open_weights parameters: "12B" context: "128K" modalities: [text] architecture_notes: > Built jointly with NVIDIA; Apache 2.0; foundation for Pixtral 12B. capabilities: [open_weights, long_context] placement_confidence: high notable: > Replaced Mistral 7B as the default open mid-weight. source_refs: [src-mistral-nemo-blog] - id: mistral-large-2 name: Mistral Large 2 vendor: Mistral AI release_date: "2024-07-24" year_bucket: 2024 branch: decoder_only secondary_tags: [open_weights, historical_frontier, long_context, eu] parents: [mistral-large-1] influences: [] openness: open_weights parameters: "123B" context: "128K" modalities: [text] architecture_notes: > MRL license; 80+ programming languages; close to Llama 3.1 405B at ~30% the parameters. capabilities: [open_weights, long_context] placement_confidence: high notable: > Mistral's strongest dense model; open-weight foundation for Pixtral Large. source_refs: [src-mistral-large-2-blog] - id: mistral-small-3 name: Mistral Small 3 vendor: Mistral AI release_date: "2025-01-30" year_bucket: 2025 branch: decoder_only secondary_tags: [open_weights, small, eu] parents: [mistral-nemo-12b] influences: [] openness: open_weights parameters: "24B" context: "32K" modalities: [text] architecture_notes: > Apache 2.0; latency-optimized dense model. capabilities: [open_weights] placement_confidence: high notable: > Mid-2020s reset of the Mistral small/medium tier. source_refs: [src-mistral-small-3-blog] # ---- Cohere ---- - id: cohere-command name: Cohere Command vendor: Cohere release_date: "2023-09-29" year_bucket: 2023 branch: decoder_only secondary_tags: [closed, enterprise] parents: [] influences: [] openness: closed parameters: "undisclosed (~52B est.)" context: "4K" modalities: [text] architecture_notes: > Cohere's first instruction-tuned production model. capabilities: [] placement_confidence: medium notable: > Anchor for the Cohere open-weights branch. source_refs: [src-cohere-command-blog] - id: command-r name: Cohere Command R vendor: Cohere release_date: "2024-03-11" year_bucket: 2024 branch: decoder_only secondary_tags: [open_weights, rag_native, multilingual, enterprise, long_context] parents: [cohere-command] influences: [] openness: open_weights parameters: "35B" context: "128K" modalities: [text] architecture_notes: > CC-BY-NC; native RAG with citation generation. capabilities: [open_weights, long_context] placement_confidence: high notable: > First open-weights model purpose-designed for RAG with grounded citations. source_refs: [src-command-r-blog] - id: command-r-plus name: Cohere Command R+ vendor: Cohere release_date: "2024-04-04" year_bucket: 2024 branch: decoder_only secondary_tags: [open_weights, rag_native, multilingual, enterprise, historical_frontier, long_context] parents: [command-r] influences: [] openness: open_weights parameters: "104B" context: "128K" modalities: [text] architecture_notes: > Dense transformer with GQA; CC-BY-NC; multi-step tool use plus JSON-schema function calling; 10 languages. capabilities: [tools, open_weights, long_context] placement_confidence: high notable: > Largest open-weights enterprise/RAG model of early 2024. source_refs: [src-command-r-plus-blog] - id: aya-23 name: Cohere Aya 23 vendor: Cohere For AI release_date: "2024-05-23" year_bucket: 2024 branch: decoder_only secondary_tags: [open_weights, multilingual] parents: [command-r] influences: [] openness: open_weights parameters: "8B, 35B" context: "8K" modalities: [text] architecture_notes: > CC-BY-NC; 23 languages. capabilities: [open_weights] placement_confidence: high notable: > Defined the multilingual open-weights frontier outside Llama. source_refs: [src-aya-23-blog] - id: aya-expanse name: Cohere Aya Expanse vendor: Cohere For AI release_date: "2024-10-23" year_bucket: 2024 branch: decoder_only secondary_tags: [open_weights, multilingual, long_context] parents: [aya-23] influences: [] openness: open_weights parameters: "8B, 32B" context: "128K" modalities: [text] architecture_notes: > CC-BY-NC; 23 languages; long-context successor to Aya 23. capabilities: [open_weights, long_context] placement_confidence: high notable: > Long-context refresh of Cohere's multilingual line. source_refs: [src-aya-expanse-docs] - id: command-a name: Cohere Command A vendor: Cohere release_date: "2025-03-11" year_bucket: 2025 branch: decoder_only secondary_tags: [open_weights, rag_native, agentic, multilingual, enterprise, long_context] parents: [command-r-plus] influences: [] openness: open_weights parameters: "111B" context: "256K" modalities: [text] architecture_notes: > Open weights for research; runs on 2x H100; 156 tok/s; 23 languages. capabilities: [tools, agentic, open_weights, long_context] placement_confidence: high notable: > Cohere's 2025 frontier; on-par with GPT-4o on agentic enterprise tasks. source_refs: [src-command-a-blog] # ---- Microsoft Phi ---- - id: phi-1 name: Phi-1 vendor: Microsoft release_date: "2023-06-21" year_bucket: 2023 branch: decoder_only secondary_tags: [open_weights, small, distilled, code_specialist] parents: [] influences: [] openness: open_weights parameters: "1.3B" context: "2048" modalities: [text] architecture_notes: > Trained on "textbook quality" filtered code data. capabilities: [open_weights] placement_confidence: high notable: > Showed quality-greater-than-scale for narrow domains; SOTA on HumanEval/MBPP for sub-2B models. source_refs: [src-phi-1-paper] - id: phi-1-5 name: Phi-1.5 vendor: Microsoft release_date: "2023-09-11" year_bucket: 2023 branch: decoder_only secondary_tags: [open_weights, small, distilled] parents: [phi-1] influences: [] openness: open_weights parameters: "1.3B" context: "2048" modalities: [text] architecture_notes: > Same architecture as Phi-1, trained on synthetic textbook data for general reasoning. capabilities: [open_weights] placement_confidence: high notable: > Matched models 5x its size on common-sense reasoning. source_refs: [src-phi-1-5-paper] - id: phi-2 name: Phi-2 vendor: Microsoft release_date: "2023-12-12" year_bucket: 2023 branch: decoder_only secondary_tags: [open_weights, small, distilled] parents: [phi-1-5] influences: [] openness: open_weights parameters: "2.7B" context: "2048" modalities: [text] architecture_notes: > Trained on 1.4T tokens of synthetic plus web data; aggressive data curation. capabilities: [open_weights] placement_confidence: high notable: > Matched or beat models up to 25x larger on complex benchmarks; flagship of the SML thesis. source_refs: [src-phi-2-blog] - id: phi-3 name: Phi-3 (mini / small / medium) vendor: Microsoft release_date: "2024-04-23" year_bucket: 2024 branch: decoder_only secondary_tags: [open_weights, small, distilled, edge, long_context] parents: [phi-2] influences: [] openness: open_weights parameters: "3.8B (mini) / 7B (small) / 14B (medium)" context: "4K / 128K" modalities: [text] architecture_notes: > Decoder-only with highly filtered web plus synthetic data; mini ships with both 4K and 128K context variants. capabilities: [open_weights, long_context] placement_confidence: high notable: > Phi-3-mini ran on a phone, comparable to GPT-3.5. source_refs: [src-phi-3-paper] - id: phi-4 name: Phi-4 vendor: Microsoft release_date: "2024-12-12" year_bucket: 2024 branch: decoder_only secondary_tags: [open_weights, distilled] parents: [phi-3] influences: [] openness: open_weights parameters: "14B" context: "16K" modalities: [text] architecture_notes: > Heavy synthetic-data training; benchmark-targeted data curation. capabilities: [open_weights] placement_confidence: high notable: > Outperformed GPT-4o on competition math; MIT-licensed weights. source_refs: [src-phi-4-paper] # ---- IBM Granite ---- - id: granite-3-0 name: IBM Granite 3.0 vendor: IBM release_date: "2024-10-21" year_bucket: 2024 branch: decoder_only secondary_tags: [open_weights, enterprise, multilingual, code_specialist] parents: [] influences: [llama-2] openness: open_weights parameters: "8B, 2B dense; 3B-A800M, 1B-A400M MoE" context: "4K" modalities: [text] architecture_notes: > Apache 2.0; trained 12T+ tokens, 12 NL languages, 116 programming languages; bundled Granite Guardian safety models. capabilities: [open_weights] placement_confidence: high notable: > IBM's reset of the Granite line into a credible enterprise open-weights family. source_refs: [src-granite-3-blog] - id: granite-3-1 name: IBM Granite 3.1 vendor: IBM release_date: "2025-01-12" year_bucket: 2025 branch: decoder_only secondary_tags: [open_weights, enterprise, long_context] parents: [granite-3-0] influences: [] openness: open_weights parameters: "2B, 8B dense; 1B-A400M, 3B-A800M MoE" context: "128K" modalities: [text] architecture_notes: > Progressive long-context training; embedding models 30M-278M added. capabilities: [open_weights, long_context] placement_confidence: high notable: > Long-context refresh of Granite. source_refs: [src-granite-3-1-repo] - id: granite-3-3 name: IBM Granite 3.3 vendor: IBM release_date: "2025-04-16" year_bucket: 2025 branch: decoder_only secondary_tags: [open_weights, enterprise, multilingual, code_specialist, long_context] parents: [granite-3-2] influences: [] openness: open_weights parameters: "2B, 8B" context: "128K" modalities: [text] architecture_notes: > Apache 2.0; refined response-length and originality controls. capabilities: [open_weights, long_context] placement_confidence: high notable: > Last dense-Transformer Granite before the 4.0 hybrid pivot. source_refs: [src-granite-3-blog] # ---- TII Falcon ---- - id: falcon-7b-40b name: Falcon 7B / 40B vendor: TII release_date: "2023-05-25" year_bucket: 2023 branch: decoder_only secondary_tags: [open_weights, uae] parents: [] influences: [] openness: open_weights parameters: "7B, 40B" context: "2K" modalities: [text] architecture_notes: > Multi-query attention plus FlashAttention; trained on RefinedWeb; relicensed Apache 2.0 May 31, 2023. capabilities: [open_weights] placement_confidence: high notable: > First Apache-2.0 non-Western frontier open-weight. source_refs: [src-falcon-card] - id: falcon-180b name: Falcon 180B vendor: TII release_date: "2023-09-06" year_bucket: 2023 branch: decoder_only secondary_tags: [open_weights, historical_frontier, uae] parents: [falcon-7b-40b] influences: [] openness: open_weights parameters: "180B" context: "2K" modalities: [text] architecture_notes: > Trained 3.5T tokens; hosting carve-out license; on par with PaLM-2 Large. capabilities: [open_weights] placement_confidence: high notable: > Largest publicly-available open-weight model of 2023 before Llama 3.1 405B. source_refs: [src-falcon-180b-blog] - id: falcon-3 name: Falcon 3 family vendor: TII release_date: "2024-12-17" year_bucket: 2024 branch: decoder_only secondary_tags: [open_weights, small, code_specialist, uae] parents: [falcon-180b] influences: [] openness: open_weights parameters: "1B, 3B, 7B, 10B" context: "32K" modalities: [text] architecture_notes: > Llama-architecture-compatible; depth up-scaling for 10B; distillation for 1B/3B. capabilities: [open_weights] placement_confidence: high notable: > Pivoted Falcon to Llama-compatible small/efficient lineup. source_refs: [src-falcon-3-blog] # ---- Qwen (Alibaba) ---- - id: qwen-7b name: Qwen-7B vendor: Alibaba release_date: "2023-08" year_bucket: 2023 branch: decoder_only secondary_tags: [open_weights, bilingual, china] parents: [] influences: [llama-2] openness: open_weights parameters: "7B" context: "32K" modalities: [text] architecture_notes: > Pretrained on 2.4T tokens; SwiGLU plus RoPE. capabilities: [open_weights] placement_confidence: high notable: > Alibaba's debut on the open-weights frontier. source_refs: [src-qwen-blog] - id: qwen-72b name: Qwen 1.8B / 14B / 72B vendor: Alibaba release_date: "2023-12" year_bucket: 2023 branch: decoder_only secondary_tags: [open_weights, china] parents: [qwen-7b] influences: [] openness: open_weights parameters: "1.8B / 14B / 72B" context: "32K" modalities: [text] architecture_notes: > Family expansion across small/medium/large tiers. capabilities: [open_weights] placement_confidence: high notable: > Qwen reaches the 72B frontier-adjacent open-weights tier. source_refs: [src-qwen-blog] - id: qwen-1-5 name: Qwen 1.5 vendor: Alibaba release_date: "2024-02" year_bucket: 2024 branch: decoder_only secondary_tags: [open_weights, china] parents: [qwen-72b] influences: [] openness: open_weights parameters: "0.5B / 1.8B / 4B / 7B / 14B / 32B / 72B / 110B" context: "32K" modalities: [text] architecture_notes: > Family expansion plus first MoE variants in the Qwen line. capabilities: [open_weights] placement_confidence: high notable: > Quietly shipped MoE in Feb 2024, predating DeepSeek-V2. source_refs: [src-qwen-1-5-blog] - id: qwen-2 name: Qwen2 vendor: Alibaba release_date: "2024-06" year_bucket: 2024 branch: decoder_only secondary_tags: [open_weights, long_context, china] parents: [qwen-1-5] influences: [] openness: open_weights parameters: "0.5B-72B" context: "128K (72B)" modalities: [text] architecture_notes: > Long-context refresh of the Qwen family. capabilities: [open_weights, long_context] placement_confidence: high notable: > Continued Alibaba's open-weights momentum. source_refs: [src-qwen-blog] - id: qwen-2-5 name: Qwen2.5 vendor: Alibaba release_date: "2024-09" year_bucket: 2024 branch: decoder_only secondary_tags: [open_weights, code_specialist, math_specialist, long_context, china] parents: [qwen-2] influences: [] openness: open_weights parameters: "0.5B / 1.5B / 3B / 7B / 14B / 32B / 72B" context: "128K" modalities: [text] architecture_notes: > Trained on up to 18T tokens; Coder/Math sibling lines. capabilities: [open_weights, long_context] placement_confidence: high notable: > Strong open-weights baseline for the Chinese frontier. source_refs: [src-qwen-2-5-blog] # ---- DeepSeek ---- - id: deepseek-llm-67b name: DeepSeek LLM 67B vendor: DeepSeek AI release_date: "2023-11-29" year_bucket: 2023 branch: decoder_only secondary_tags: [open_weights, china] parents: [] influences: [llama-2] openness: open_weights parameters: "67B" context: "4K" modalities: [text] architecture_notes: > Dense transformer with Grouped-Query Attention; trained on 2T English and Chinese tokens. capabilities: [open_weights] placement_confidence: high notable: > DeepSeek's debut; outperformed Llama-2-70B on code and math. source_refs: [src-deepseek-llm-repo] - id: deepseek-coder name: DeepSeek-Coder vendor: DeepSeek AI release_date: "2023-11" year_bucket: 2023 branch: decoder_only secondary_tags: [open_weights, code_specialist, china] parents: [deepseek-llm-67b] influences: [] openness: open_weights parameters: "1.3B / 6.7B / 33B" context: "16K" modalities: [text] architecture_notes: > Trained on 2T tokens with 87% code; FIM objective. capabilities: [open_weights] placement_confidence: high notable: > First DeepSeek code-specialized model. source_refs: [src-deepseek-coder-page] - id: deepseek-math name: DeepSeekMath 7B vendor: DeepSeek AI release_date: "2024-02" year_bucket: 2024 branch: decoder_only secondary_tags: [open_weights, math_specialist, china] parents: [deepseek-coder] influences: [] openness: open_weights parameters: "7B" context: "4K" modalities: [text] architecture_notes: > Continued from DeepSeek-Coder-Base-v1.5 with 120B math tokens; introduced GRPO. capabilities: [open_weights] placement_confidence: high notable: > Pioneered Group Relative Policy Optimization (GRPO), later used in R1. source_refs: [src-deepseekmath-paper] # ---- Yi ---- - id: yi-34b name: Yi-34B vendor: 01.AI release_date: "2023-11" year_bucket: 2023 branch: decoder_only secondary_tags: [open_weights, bilingual, china] parents: [] influences: [llama-2] openness: open_weights parameters: "6B, 34B" context: "4K (extended to 200K variants)" modalities: [text] architecture_notes: > Bilingual (English plus Chinese); long-context variants up to 200K. capabilities: [open_weights] placement_confidence: high notable: > 01.AI's open-weights debut. source_refs: [src-yi-collection] - id: yi-1-5 name: Yi-1.5 vendor: 01.AI release_date: "2024-05-13" year_bucket: 2024 branch: decoder_only secondary_tags: [open_weights, china] parents: [yi-34b] influences: [] openness: open_weights parameters: "6B / 9B / 34B" context: "4K (extendable)" modalities: [text] architecture_notes: > Continued pretraining on 500B tokens plus 3M SFT samples. capabilities: [open_weights] placement_confidence: high notable: > Refined the Yi-34B baseline. source_refs: [src-yi-1-5-repo] # ---- ERNIE / Baidu ---- - id: ernie-4 name: ERNIE 4.0 vendor: Baidu release_date: "2023-10" year_bucket: 2023 branch: decoder_only secondary_tags: [closed, multilingual, china] parents: [] influences: [] openness: closed parameters: "undisclosed" context: "8K" modalities: [text] architecture_notes: > Baidu's flagship dense decoder. capabilities: [] placement_confidence: high notable: > China's GPT-4 peer at the time. source_refs: [src-ernie-baike] # ---- xAI Grok ---- - id: grok-1 name: Grok-1 vendor: xAI release_date: "2023-11-04" year_bucket: 2023 branch: decoder_only secondary_tags: [open_weights, mixture_of_experts] parents: [] influences: [] openness: open_weights parameters: "314B (MoE, 25% active)" context: "8K" modalities: [text] architecture_notes: > Open-sourced March 2024 under Apache 2.0. capabilities: [open_weights] placement_confidence: high notable: > xAI's open-weights debut. source_refs: [src-grok-os-blog] - id: grok-1-5 name: Grok-1.5 vendor: xAI release_date: "2024-03" year_bucket: 2024 branch: decoder_only secondary_tags: [closed, long_context] parents: [grok-1] influences: [] openness: closed parameters: "undisclosed" context: "128K" modalities: [text] architecture_notes: > Closed iteration with long context. capabilities: [long_context] placement_confidence: high notable: > Bridge to the multimodal Grok-1.5V. source_refs: [src-grok-history-page] - id: grok-2 name: Grok-2 vendor: xAI release_date: "2024-08" year_bucket: 2024 branch: decoder_only secondary_tags: [closed, multimodal, tool_using] parents: [grok-1-5] influences: [] openness: closed parameters: "undisclosed" context: "128K" modalities: [text, image] architecture_notes: > Closed flagship with vision and tool use. capabilities: [tools, multimodal] placement_confidence: high notable: > Last pre-reasoning xAI flagship. source_refs: [src-grok-history-page] # ---- AI21 Jamba (SSM-hybrid as secondary tag) ---- - id: jurassic-2 name: AI21 Jurassic-2 vendor: AI21 Labs release_date: "2023-03-09" year_bucket: 2023 branch: decoder_only secondary_tags: [closed, multilingual] parents: [] influences: [] openness: closed parameters: "Light 7B, Mid 17B, Ultra 60B" context: "8K" modalities: [text] architecture_notes: > Closed weights via API; included as Jamba lineage anchor. capabilities: [] placement_confidence: medium notable: > Pre-Jamba dense lineage. source_refs: [src-jurassic-2-blog] - id: jamba name: AI21 Jamba v0.1 vendor: AI21 Labs release_date: "2024-03-28" year_bucket: 2024 branch: decoder_only secondary_tags: [open_weights, ssm_hybrid, mamba, mixture_of_experts, long_context] parents: [jurassic-2] influences: [] openness: open_weights parameters: "52B total / 12B active" context: "256K" modalities: [text] architecture_notes: > Transformer/Mamba hybrid blocks (1 attn : 7 mamba : MoE); 16 experts top-2; Apache 2.0. capabilities: [open_weights, long_context] placement_confidence: high notable: > First production-grade open SSM-Transformer hybrid. source_refs: [src-jamba-blog] - id: jamba-1-5 name: AI21 Jamba 1.5 vendor: AI21 Labs release_date: "2024-08-22" year_bucket: 2024 branch: decoder_only secondary_tags: [open_weights, ssm_hybrid, mamba, mixture_of_experts, long_context] parents: [jamba] influences: [] openness: open_weights parameters: "Mini 52B/12B active; Large 398B/94B active" context: "256K" modalities: [text] architecture_notes: > Jamba Open Model License; Mini and Large size split. capabilities: [open_weights, long_context] placement_confidence: high notable: > Scaled Jamba SSM-hybrid into a frontier-class open model. source_refs: [src-jamba-1-5-blog] - id: jamba-2 name: AI21 Jamba 2 vendor: AI21 Labs release_date: "2026-01-08" year_bucket: 2026 branch: decoder_only secondary_tags: [open_weights, ssm_hybrid, mamba, mixture_of_experts, long_context] parents: [jamba-1-5] influences: [] openness: open_weights parameters: "Large 398B/94B; Mini 52B/12B; 3B" context: "256K" modalities: [text] architecture_notes: > Jamba 2 family adds a 3B edge variant; positioned for enterprise reliability. capabilities: [open_weights, long_context] placement_confidence: high notable: > Latest evolution of the Western SSM-hybrid branch. source_refs: [src-jamba-2-page] # ---- GLM closed family (decoder-only) ---- - id: glm-4 name: GLM-4 vendor: Z.AI (Zhipu) release_date: "2024-01" year_bucket: 2024 branch: decoder_only secondary_tags: [closed, china, long_context] parents: [] influences: [] openness: closed parameters: "Undisclosed flagship" context: "128K" modalities: [text] architecture_notes: > Zhipu's flagship dense model. capabilities: [long_context] placement_confidence: medium notable: > Anchor for the GLM-Z1 reasoning fork and GLM-5 MoE family. source_refs: [src-glm-page] - id: glm-4-32b-0414 name: GLM-4-32B-0414 vendor: Z.AI (Zhipu) release_date: "2025-04" year_bucket: 2025 branch: decoder_only secondary_tags: [open_weights, china, long_context] parents: [glm-4] influences: [] openness: open_weights parameters: "32B" context: "128K" modalities: [text] architecture_notes: > Pretrained on 15T high-quality data including reasoning synthetic data. capabilities: [open_weights, long_context] placement_confidence: high notable: > Zhipu's open-weights baseline for the GLM-Z1 reasoning fork. source_refs: [src-glm-4-repo] # ---- Kimi base ---- - id: kimi-k1 name: Kimi K1 vendor: Moonshot AI release_date: "2024-03" year_bucket: 2024 branch: decoder_only secondary_tags: [closed, china, long_context] parents: [] influences: [] openness: closed parameters: "Undisclosed" context: "200K (chat product)" modalities: [text] architecture_notes: > Initially marketed as long-context Chinese chat product. capabilities: [long_context] placement_confidence: medium notable: > Origin of the Kimi line. source_refs: [src-kimi-k1-page] # ============================================================ # MIXTURE-OF-EXPERTS # ============================================================ - id: switch-transformer name: Switch Transformer vendor: Google release_date: "2021-01-11" year_bucket: 2021 branch: mixture_of_experts secondary_tags: [open_source, foundational] parents: [t5] influences: [] openness: open_source parameters: "Up to 1.6T total (sparsely activated)" context: "512" modalities: [text] architecture_notes: > Simplified MoE routing — each token sent to a single expert; bf16 sparse training; built on T5. capabilities: [] placement_confidence: high notable: > First trillion-parameter LM; defined the modern MoE pattern at scale. source_refs: [src-switch-transformer-paper] - id: glam name: GLaM vendor: Google release_date: "2021-12-09" year_bucket: 2021 branch: mixture_of_experts secondary_tags: [closed] parents: [] influences: [switch-transformer] openness: closed parameters: "1.2T total / 97B active per token" context: "1024" modalities: [text] architecture_notes: > Decoder-only sparse MoE — 64 experts per MoE layer, top-2 gating; 7x larger than GPT-3 with 1/3 the training energy. capabilities: [] placement_confidence: high notable: > Demonstrated MoE works for decoder-only generation. source_refs: [src-glam-paper] - id: st-moe name: ST-MoE vendor: Google release_date: "2022-02-17" year_bucket: 2022 branch: mixture_of_experts secondary_tags: [closed] parents: [switch-transformer] influences: [] openness: closed parameters: "269B total (32B compute-equivalent)" context: "1024" modalities: [text] architecture_notes: > Stabilization techniques (router z-loss, update precision rules) plus sparse-specific fine-tuning; encoder-decoder backbone. capabilities: [] placement_confidence: high notable: > First sparse model SOTA on transfer learning; informed downstream MoE in Gemini. source_refs: [src-st-moe-paper] - id: mixtral-8x7b name: Mixtral 8x7B vendor: Mistral AI release_date: "2023-12-11" year_bucket: 2023 branch: mixture_of_experts secondary_tags: [open_weights, eu, long_context] parents: [mistral-7b] influences: [glam] openness: open_weights parameters: "46.7B total / 12.9B active" context: "32K" modalities: [text] architecture_notes: > Sparse MoE: 8 experts, top-2 routing per token; Apache 2.0. capabilities: [open_weights] placement_confidence: high notable: > First widely-deployed open-weights MoE; defined the open-MoE branch. source_refs: [src-mixtral-blog] - id: mixtral-8x22b name: Mixtral 8x22B vendor: Mistral AI release_date: "2024-04-10" year_bucket: 2024 branch: mixture_of_experts secondary_tags: [open_weights, multilingual, eu, long_context] parents: [mixtral-8x7b] influences: [] openness: open_weights parameters: "141B total / 39B active" context: "64K" modalities: [text] architecture_notes: > Sparse MoE; Apache 2.0; multilingual EN/FR/IT/DE/ES. capabilities: [open_weights, long_context] placement_confidence: high notable: > Largest fully Apache-2.0 MoE before DBRX/Arctic. source_refs: [src-mixtral-8x22b-blog] - id: dbrx name: DBRX vendor: Databricks release_date: "2024-03-27" year_bucket: 2024 branch: mixture_of_experts secondary_tags: [open_weights, enterprise] parents: [mixtral-8x7b] influences: [] openness: open_weights parameters: "132B total / 36B active" context: "32K" modalities: [text] architecture_notes: > Fine-grained MoE: 16 experts, top-4 (vs Mixtral's 8/top-2); ~2x faster inference than Llama 2 70B; trained for ~$10M. capabilities: [open_weights] placement_confidence: high notable: > Pushed fine-grained MoE into the open; influenced Llama 4 routing design. source_refs: [src-dbrx-blog] - id: snowflake-arctic name: Snowflake Arctic vendor: Snowflake release_date: "2024-04-24" year_bucket: 2024 branch: mixture_of_experts secondary_tags: [open_weights, enterprise] parents: [dbrx, mixtral-8x22b] influences: [] openness: open_weights parameters: "480B total / 17B active" context: "4K (initial)" modalities: [text] architecture_notes: > Dense-MoE Hybrid: 10B dense transformer plus residual 128 x 3.66B MoE MLP; top-2 gating; Apache 2.0; trained for ~$2M (17x less than Llama 3 70B). capabilities: [open_weights] placement_confidence: high notable: > Largest open-weights MoE of 2024 by total params. source_refs: [src-arctic-blog] - id: phi-3-5-moe name: Phi-3.5 MoE vendor: Microsoft release_date: "2024-08-21" year_bucket: 2024 branch: mixture_of_experts secondary_tags: [open_weights, small, distilled, long_context] parents: [phi-3] influences: [mixtral-8x7b] openness: open_weights parameters: "41.9B total / 6.6B active (16 experts)" context: "128K" modalities: [text] architecture_notes: > First Phi-family MoE; GRIN training method; MIT license. capabilities: [open_weights, long_context] placement_confidence: high notable: > Brought MoE into the small-model branch. source_refs: [src-phi-3-5-card] - id: deepseek-v2 name: DeepSeek-V2 vendor: DeepSeek AI release_date: "2024-05-06" year_bucket: 2024 branch: mixture_of_experts secondary_tags: [open_weights, china, long_context] parents: [deepseek-llm-67b] influences: [] openness: open_weights parameters: "236B total / 21B active" context: "128K" modalities: [text] architecture_notes: > First DeepSeekMoE; introduced Multi-head Latent Attention (MLA); 93% KV-cache reduction vs 67B. capabilities: [open_weights, long_context] placement_confidence: high notable: > Inflection point — MLA plus DeepSeekMoE rewired Chinese training economics. source_refs: [src-deepseek-v2-paper] - id: deepseek-coder-v2 name: DeepSeek-Coder-V2 vendor: DeepSeek AI release_date: "2024-06" year_bucket: 2024 branch: mixture_of_experts secondary_tags: [open_weights, code_specialist, china, long_context] parents: [deepseek-v2] influences: [] openness: open_weights parameters: "236B/21B (Std); 16B/2.4B (Lite)" context: "128K" modalities: [text] architecture_notes: > Continued pretraining on +6T tokens; 338 programming languages. capabilities: [open_weights, long_context] placement_confidence: high notable: > Extended DeepSeek-V2 to coding workloads. source_refs: [src-deepseek-coder-v2-page] - id: deepseek-v3 name: DeepSeek-V3 vendor: DeepSeek AI release_date: "2024-12-26" year_bucket: 2024 branch: mixture_of_experts secondary_tags: [open_weights, historical_frontier, china, long_context] parents: [deepseek-v2] influences: [] openness: open_weights parameters: "671B total / 37B active" context: "128K" modalities: [text] architecture_notes: > MoE with MLA; FP8 mixed-precision training on ~2,048 H800 GPUs in under 2 months for ~$5.6M. capabilities: [open_weights, long_context] placement_confidence: high notable: > The cost-shock release that reframed economics of frontier training. source_refs: [src-deepseek-v3-paper] - id: deepseek-v3-0324 name: DeepSeek-V3-0324 vendor: DeepSeek AI release_date: "2025-03-24" year_bucket: 2025 branch: mixture_of_experts secondary_tags: [open_weights, china, long_context] parents: [deepseek-v3] influences: [deepseek-r1] openness: open_weights parameters: "671B total / 37B active" context: "128K" modalities: [text] architecture_notes: > Reasoning-distilled refresh; +5.3 MMLU-Pro, +9.3 GPQA, +19.8 AIME 2025 vs V3. capabilities: [open_weights, long_context] placement_confidence: high notable: > First DeepSeek refresh after R1 propagated reasoning traits back into the base. source_refs: [src-deepseek-version-history] - id: deepseek-v3-2 name: DeepSeek-V3.2 vendor: DeepSeek AI release_date: "2025-12" year_bucket: 2025 branch: mixture_of_experts secondary_tags: [open_weights, china, long_context] parents: [deepseek-v3-0324] influences: [] openness: open_weights parameters: "671B total / 37B active" context: "128K" modalities: [text] architecture_notes: > Default DeepSeek model immediately prior to V4; introduced DeepSeek Sparse Attention (DSA). capabilities: [open_weights, long_context] placement_confidence: medium notable: > Introduced DSA, the first cross-vendor architectural pattern to propagate within the Chinese frontier. source_refs: [src-deepseek-version-history] - id: deepseek-v4-pro name: DeepSeek-V4 Pro vendor: DeepSeek AI release_date: "2026-04-24" year_bucket: 2026 branch: mixture_of_experts secondary_tags: [open_weights, frontier, china, ultra_long_context, agentic] parents: [deepseek-v3-2] influences: [] openness: open_weights parameters: "1.6T total / 49B active" context: "1M" modalities: [text] architecture_notes: > Hybrid attention with multi-head Conditional Attention, Engram Conditional Memory, DSA sparse attention; 73% per-token FLOP reduction, 90% KV-cache reduction. Pretrained on 32T+ tokens; 384,000 maximum output tokens. Served as a preview build from 2026-04-24 until 2026-08-12, when DeepSeek swapped the deepseek-v4-pro endpoint to the 0813 general-availability build with no blog post and no announcement. capabilities: [tools, agentic, open_weights, long_context] placement_confidence: high notable: > First Chinese open-weights model to match Western frontier on aggregate benchmarks (BenchLM 87). Updated 2026-08-12 for the 0813 GA build, which is where the openness label starts to strain: DeepSeek reports large gains over the April preview (Terminal Bench 2.1 72.1 to 87.9, DeepSWE 12.8 to 62.7, CyberGym 52.7 to 83.3, AA Intelligence Index 45 to 53), but the Hugging Face repositories still hosted the April preview weights after the swap. For a period, every independent evaluation of "DeepSeek V4 Pro" was measuring a build no one could download, and every downloadable build was one no one was serving. A price change was also announced for 2026-08-16 replacing the flat $0.435 input / $0.87 output rate with peak and off-peak tiers reported at $1.32 / $3.96 peak and half that off-peak. source_refs: [src-deepseek-v4-news, src-deepseek-v4-pro-0813] - id: deepseek-v4-flash name: DeepSeek-V4 Flash vendor: DeepSeek AI release_date: "2026-04-24" year_bucket: 2026 branch: mixture_of_experts secondary_tags: [open_weights, china, ultra_long_context, agentic] parents: [deepseek-v3-2] influences: [] openness: open_weights parameters: "284B total / 13B active" context: "1M" modalities: [text] architecture_notes: > Cost-efficient sibling to V4 Pro; same 1M context. capabilities: [tools, agentic, open_weights, long_context] placement_confidence: high notable: > Cost-efficient companion to V4 Pro at consumer-deployable scale. source_refs: [src-deepseek-v4-flash-news] - id: qwen-3 name: Qwen3 (235B-A22B / 30B-A3B) vendor: Alibaba release_date: "2025-04" year_bucket: 2025 branch: mixture_of_experts secondary_tags: [open_weights, hybrid_reasoning, frontier, china, long_context] parents: [qwen-2-5] influences: [deepseek-v3, deepseek-r1] openness: open_weights parameters: "235B/22B (flagship MoE); 30B/3B; dense 0.6B-32B" context: "128K" modalities: [text] architecture_notes: > Hybrid thinking modes — explicit reasoning toggle in API/prompt. capabilities: [open_weights, long_context] placement_confidence: high notable: > First major Western/Eastern model with a hybrid thinking-mode toggle. source_refs: [src-qwen-3-blog] - id: kimi-k2 name: Kimi K2 vendor: Moonshot AI release_date: "2025-07" year_bucket: 2025 branch: mixture_of_experts secondary_tags: [open_weights, agentic, china, long_context] parents: [kimi-k1-5] influences: [deepseek-v3] openness: open_weights parameters: "1T total / 32B active" context: "128K" modalities: [text] architecture_notes: > Trillion-parameter MoE; agentic-coding-first. capabilities: [tools, agentic, open_weights, long_context] placement_confidence: high notable: > First open-weights trillion-parameter MoE. source_refs: [src-kimi-wikipedia] - id: kimi-k2-6 name: Kimi K2.6 vendor: Moonshot AI release_date: "2026-04-20" year_bucket: 2026 branch: mixture_of_experts secondary_tags: [open_weights, agentic, multi_agent_native, frontier, china, multimodal, long_context] parents: [kimi-k2] influences: [] openness: open_weights parameters: "1T total / 32B active (384 routed experts: 8 active + 1 shared)" context: "262K" modalities: [text, image] architecture_notes: > MLA plus SwiGLU; MoonViT 400M vision encoder; automatic context compression for 12-hour autonomous runs; native 300-sub-agent swarm. capabilities: [tools, agentic, multi_agent_native, multimodal, open_weights, long_context] placement_confidence: high notable: > Native 300-sub-agent swarm orchestration; 58.6% SWE-Bench Pro. source_refs: [src-kimi-k2-6-blog] - id: glm-5 name: GLM-5 vendor: Z.AI (Zhipu) release_date: "2026-02" year_bucket: 2026 branch: mixture_of_experts secondary_tags: [open_weights, frontier, china, huawei_trained, long_context] parents: [glm-4-32b-0414] influences: [deepseek-v3] openness: open_weights parameters: "744B total / 40B active" context: "200K" modalities: [text] architecture_notes: > Trained on 28.5T tokens entirely on Huawei Ascend 910C; 128 MoE experts; integrates DeepSeek Sparse Attention; Apache 2.0. capabilities: [open_weights, long_context] placement_confidence: high notable: > First frontier-tier Chinese model trained without NVIDIA hardware. source_refs: [src-glm-5-repo] - id: hunyuan-large name: Hunyuan-Large vendor: Tencent release_date: "2024-11" year_bucket: 2024 branch: mixture_of_experts secondary_tags: [open_weights, china, long_context] parents: [] influences: [] openness: open_weights parameters: "389B total / 52B active" context: "256K" modalities: [text] architecture_notes: > Largest open-source Transformer MoE at release. capabilities: [open_weights, long_context] placement_confidence: high notable: > Tencent's open-weights frontier debut. source_refs: [src-hunyuan-large-repo] - id: hunyuan-turbos name: Hunyuan-TurboS vendor: Tencent release_date: "2025-03" year_bucket: 2025 branch: mixture_of_experts secondary_tags: [closed, ssm_hybrid, mamba, china, long_context] parents: [hunyuan-large] influences: [] openness: closed parameters: "560B total / 56B active" context: "256K" modalities: [text] architecture_notes: > First ultra-large Hybrid Transformer-Mamba MoE; adaptive CoT switch between fast/deep. capabilities: [long_context] placement_confidence: high notable: > First ultra-large SSM-hybrid MoE. source_refs: [src-hunyuan-turbos-repo] - id: hunyuan-2 name: Hunyuan 2.0 vendor: Tencent release_date: "2025-12" year_bucket: 2025 branch: mixture_of_experts secondary_tags: [closed, china, long_context] parents: [hunyuan-turbos] influences: [] openness: closed parameters: "406B total / 32B active" context: "256K" modalities: [text] architecture_notes: > Tencent's late-2025 MoE refresh. capabilities: [long_context] placement_confidence: medium notable: > Continued the Hunyuan MoE line. source_refs: [src-hunyuan-2-news] - id: minimax-text-01 name: MiniMax-Text-01 vendor: MiniMax release_date: "2025-01-15" year_bucket: 2025 branch: mixture_of_experts secondary_tags: [open_weights, china, ultra_long_context] parents: [] influences: [] openness: open_weights parameters: "456B total / 45.9B active" context: "1M (4M extrapolation)" modalities: [text] architecture_notes: > Hybrid Lightning Attention plus softmax (7:1 ratio); 32 experts top-2 routing. capabilities: [open_weights, long_context] placement_confidence: high notable: > Pioneered Lightning Attention at frontier scale. source_refs: [src-minimax-paper] - id: gpt-oss-120b name: gpt-oss-120b vendor: OpenAI release_date: "2025-08-05" year_bucket: 2025 branch: mixture_of_experts secondary_tags: [open_weights, agentic, long_context] parents: [] influences: [gpt-3, o4-mini] openness: open_weights parameters: "116.8B total / 5.1B active per token" context: "128K" modalities: [text] architecture_notes: > Autoregressive MoE transformer. Native MXFP4 quantization in MoE layer. Configurable reasoning-effort. Trained with RL informed by frontier models like o3. capabilities: [tools, agentic, open_weights, long_context] placement_confidence: high notable: > OpenAI's first open-weight release since GPT-2 (2019). Apache 2.0. Near-parity with o4-mini on a single 80GB GPU. source_refs: [src-gpt-oss-paper] - id: gpt-oss-20b name: gpt-oss-20b vendor: OpenAI release_date: "2025-08-05" year_bucket: 2025 branch: mixture_of_experts secondary_tags: [open_weights, small, edge, agentic, long_context] parents: [gpt-oss-120b] influences: [o3-mini] openness: open_weights parameters: "20.9B total / 3.6B active per token" context: "128K" modalities: [text] architecture_notes: > Same MoE family as 120b, smaller. Runs on 16GB consumer hardware. capabilities: [tools, agentic, open_weights, long_context] placement_confidence: high notable: > First time OpenAI shipped a frontier-style reasoning model that fits on a laptop GPU. source_refs: [src-gpt-oss-blog] - id: llama-4-scout name: Llama 4 Scout vendor: Meta AI release_date: "2025-04-05" year_bucket: 2025 branch: mixture_of_experts secondary_tags: [open_weights, multimodal, frontier, ultra_long_context] parents: [llama-3-3-70b] influences: [mixtral-8x7b] openness: open_weights parameters: "109B total / 17B active" context: "10M (claimed)" modalities: [text, image, video] architecture_notes: > 16 experts, top-1-style routing; natively multimodal (early-fusion); fits on a single H100; trained on ~40T tokens, 200 languages. capabilities: [tools, multimodal, open_weights, long_context] placement_confidence: high notable: > First open-weights model with a 10M-token context window; Meta's pivot to MoE. source_refs: [src-llama-4-blog] - id: llama-4-maverick name: Llama 4 Maverick vendor: Meta AI release_date: "2025-04-05" year_bucket: 2025 branch: mixture_of_experts secondary_tags: [open_weights, multimodal, frontier, long_context] parents: [llama-3-3-70b] influences: [mixtral-8x22b, dbrx] openness: open_weights parameters: "400B total / 17B active" context: "1M" modalities: [text, image, video] architecture_notes: > 128 experts; natively multimodal; benchmarked vs. GPT-4o and Gemini 2.0 Flash; requires H100 DGX. capabilities: [tools, multimodal, open_weights, long_context] placement_confidence: high notable: > Largest open-weights MoE released by Meta; replaced dense flagship strategy. source_refs: [src-llama-4-blog] - id: mistral-large-3 name: Mistral Large 3 vendor: Mistral AI release_date: "2025-12-08" year_bucket: 2025 branch: mixture_of_experts secondary_tags: [open_weights, multimodal, frontier, eu, long_context] parents: [mistral-large-2, pixtral-large] influences: [llama-4-maverick, deepseek-v3] openness: open_weights parameters: "675B total / 41B active" context: "256K" modalities: [text, image] architecture_notes: > Granular MoE; Apache 2.0 (unrestricted commercial). capabilities: [tools, multimodal, open_weights, long_context] placement_confidence: high notable: > Mistral's pivot to large MoE; matched closed frontier under Apache 2.0. source_refs: [src-mistral-large-3-docs] - id: mistral-small-4 name: Mistral Small 4 vendor: Mistral AI release_date: "2026-03-16" year_bucket: 2026 branch: mixture_of_experts secondary_tags: [open_weights, hybrid_reasoning, code_specialist, eu, long_context] parents: [mistral-small-3, magistral] influences: [mistral-large-3] openness: open_weights parameters: "119B total / 6B active" context: "256K" modalities: [text, image] architecture_notes: > Hybrid model unifying instruct/reasoning/coding with configurable reasoning effort; MoE. capabilities: [tools, multimodal, open_weights, long_context] placement_confidence: high notable: > First widely-released hybrid mode-switching open-weight in the West. source_refs: [src-mistral-small-4-blog] - id: granite-4-0 name: IBM Granite 4.0 vendor: IBM release_date: "2025-10-02" year_bucket: 2025 branch: mixture_of_experts secondary_tags: [open_weights, ssm_hybrid, mamba, enterprise, long_context] parents: [granite-3-3] influences: [jamba, falcon-mamba-7b] openness: open_weights parameters: "350M to 32B (e.g., H-Small 32B/9B active)" context: "128K (eval) / 512K (train)" modalities: [text] architecture_notes: > Hybrid Mamba-2/Transformer with ~9:1 ratio of SSM:attention layers; MoE in select sizes; first open model family ISO 42001 certified; Apache 2.0. capabilities: [open_weights, long_context] placement_confidence: high notable: > 70% lower memory and 2x faster inference than dense peers. source_refs: [src-granite-4-blog] # SSM-only model (placed on decoder_only with ssm_hybrid tag) - id: falcon-mamba-7b name: Falcon Mamba 7B vendor: TII release_date: "2024-08-12" year_bucket: 2024 branch: decoder_only secondary_tags: [open_weights, ssm_hybrid, mamba, attention_free, uae] parents: [] influences: [] openness: open_weights parameters: "7B" context: "unbounded (constant-mem)" modalities: [text] architecture_notes: > Pure SSM (no attention); trained 5.5T tokens; constant inference throughput vs. context length. capabilities: [open_weights] placement_confidence: medium notable: > First top-ranked pure SSM in the open. source_refs: [src-falcon-mamba-blog] # ============================================================ # MULTIMODAL # ============================================================ - id: flamingo name: Flamingo vendor: DeepMind release_date: "2022-04-29" year_bucket: 2022 branch: multimodal secondary_tags: [closed, foundational] parents: [chinchilla] influences: [] openness: closed parameters: "3B / 9B / 80B" context: "Variable" modalities: [text, image, video] architecture_notes: > Frozen Chinchilla LM plus frozen vision encoder plus Perceiver Resampler plus gated cross-attention; handles arbitrarily interleaved image/video/text. capabilities: [multimodal] placement_confidence: high notable: > Defined the few-shot VLM pattern; SOTA on 16 vision-language benchmarks; design precursor to Gemini's native multimodality. source_refs: [src-flamingo-paper] - id: kosmos-1 name: Kosmos-1 vendor: Microsoft release_date: "2023-02-27" year_bucket: 2023 branch: multimodal secondary_tags: [research_only] parents: [] influences: [flamingo] openness: research_only parameters: "1.6B" context: "2048" modalities: [text, image] architecture_notes: > MLLM trained from scratch on web-scale interleaved text/image corpora; Magneto transformer backbone. capabilities: [multimodal] placement_confidence: high notable: > Early native MLLM from Microsoft. source_refs: [src-kosmos-1-paper] - id: kosmos-2 name: Kosmos-2 vendor: Microsoft release_date: "2023-06-26" year_bucket: 2023 branch: multimodal secondary_tags: [research_only] parents: [kosmos-1] influences: [] openness: research_only parameters: "1.6B" context: "2048" modalities: [text, image] architecture_notes: > Adds visual grounding via Markdown-style location tokens trained on GRIT corpus. capabilities: [multimodal] placement_confidence: high notable: > Pioneered bbox-as-tokens grounding pattern that influenced later VLMs. source_refs: [src-kosmos-2-paper] - id: gpt-4v name: GPT-4V (Vision) vendor: OpenAI release_date: "2023-09-25" year_bucket: 2023 branch: multimodal secondary_tags: [closed] parents: [gpt-4] influences: [] openness: closed parameters: "undisclosed" context: "8K-128K" modalities: [text, image] architecture_notes: > GPT-4 with vision encoder bolted on; not a native-multimodal training run. capabilities: [tools, multimodal] placement_confidence: medium notable: > First widely-deployed vision-capable LLM in production. source_refs: [src-gpt-4v-blog] - id: gpt-4o name: GPT-4o vendor: OpenAI release_date: "2024-05-13" year_bucket: 2024 branch: multimodal secondary_tags: [closed, historical_frontier, long_context] parents: [gpt-4-turbo] influences: [gpt-4v] openness: closed parameters: "undisclosed" context: "128K" modalities: [text, image, audio] architecture_notes: > First flagship trained natively across text/image/audio in a single model. Enabled real-time voice with low latency. capabilities: [tools, multimodal, long_context] placement_confidence: high notable: > The "omni" moment. Free for ChatGPT users at launch. source_refs: [src-gpt-4o-blog] - id: gpt-4o-mini name: GPT-4o mini vendor: OpenAI release_date: "2024-07-18" year_bucket: 2024 branch: multimodal secondary_tags: [closed, small, distilled, long_context] parents: [gpt-4o] influences: [] openness: closed parameters: "undisclosed (small)" context: "128K" modalities: [text, image] architecture_notes: > Smaller GPT-4o-class model. Replaced GPT-3.5 Turbo as the default cheap-and-fast option. capabilities: [tools, multimodal, long_context] placement_confidence: high notable: > Cheaper than GPT-3.5 Turbo with significantly better quality. source_refs: [src-gpt-4o-mini-blog] - id: gpt-4-1 name: GPT-4.1 vendor: OpenAI release_date: "2025-04-14" year_bucket: 2025 branch: multimodal secondary_tags: [closed, code_specialist, ultra_long_context] parents: [gpt-4o] influences: [gpt-4-5] openness: closed parameters: "undisclosed" context: "1M" modalities: [text, image] architecture_notes: > Coding/instruction-following iteration of the GPT-4o lineage; mini and nano variants released same day. capabilities: [tools, multimodal, long_context] placement_confidence: high notable: > First OpenAI flagship at 1M-token context. source_refs: [src-gpt-4-1-blog] - id: gpt-image-1 name: gpt-image-1 vendor: OpenAI release_date: "2025-04-23" year_bucket: 2025 branch: multimodal secondary_tags: [closed] parents: [gpt-4o] influences: [] openness: closed parameters: "undisclosed" context: "n/a" modalities: [text, image] architecture_notes: > Native image-generation model exposed as standalone API in addition to GPT-4o's image generation. capabilities: [multimodal] placement_confidence: medium notable: > DALL·E branding effectively retired in favor of this family. source_refs: [src-gpt-image-1-docs] - id: sora name: Sora (Sora Turbo) vendor: OpenAI release_date: "2024-12-09" year_bucket: 2024 branch: multimodal secondary_tags: [closed, video, diffusion_transformer] parents: [] influences: [gpt-4] openness: closed parameters: "undisclosed" context: "n/a (video)" modalities: [text, video, image] architecture_notes: > Diffusion-transformer (DiT) for video. Separate architectural lineage from the GPT decoder-only family. capabilities: [multimodal] placement_confidence: high notable: > OpenAI's first major video-generation model. Architecturally a sibling, not a descendant, of the GPT line. source_refs: [src-sora-blog] display_in_tree: false - id: sora-2 name: Sora 2 vendor: OpenAI release_date: "2025-09-30" year_bucket: 2025 branch: multimodal secondary_tags: [closed, video, diffusion_transformer] parents: [sora] influences: [] openness: closed parameters: "undisclosed" context: "n/a (video)" modalities: [text, video, image, audio] architecture_notes: > More physically accurate, controllable. Added cameo feature. capabilities: [multimodal] placement_confidence: high notable: > Closest OpenAI has come to a Veo-class video model. source_refs: [src-sora-2-blog] display_in_tree: false - id: gemini-1 name: Gemini 1.0 vendor: Google + DeepMind release_date: "2023-12-06" year_bucket: 2023 branch: multimodal secondary_tags: [closed, foundational] parents: [palm-2] influences: [chinchilla, flamingo] openness: closed parameters: "Nano (1.8B/3.25B) / Pro / Ultra (undisclosed)" context: "32K" modalities: [text, image, audio, video] architecture_notes: > Decoder-only transformer trained natively multimodal from scratch; Ultra/Pro/Nano tiers. capabilities: [multimodal] placement_confidence: high notable: > First model to outperform human experts on MMLU (90.0% Ultra). Marked the merger of Google Brain plus DeepMind into one model line. source_refs: [src-gemini-1-paper] - id: gemini-1-5 name: Gemini 1.5 (Pro / Flash) vendor: Google + DeepMind release_date: "2024-02-15" year_bucket: 2024 branch: multimodal secondary_tags: [closed, mixture_of_experts, ultra_long_context] parents: [gemini-1] influences: [st-moe] openness: closed parameters: "Undisclosed; Pro and Flash tiers" context: "1M (up to 2M experimentally)" modalities: [text, image, audio, video] architecture_notes: > Sparse MoE transformer with native multimodality; Flash distilled from Pro for low-latency. capabilities: [multimodal, long_context] placement_confidence: high notable: > First widely deployed 1M+ token context window. source_refs: [src-gemini-1-5-paper] - id: gemini-2 name: Gemini 2.0 (Flash / Pro / Flash-Lite) vendor: Google + DeepMind release_date: "2024-12-11" year_bucket: 2024 branch: multimodal secondary_tags: [closed, mixture_of_experts, agentic, long_context] parents: [gemini-1-5] influences: [] openness: closed parameters: "Undisclosed (Flash / Pro / Flash-Lite tiers)" context: "1M+" modalities: [text, image, audio, video] architecture_notes: > Native multimodal output (image plus audio); native tool use (Search, code exec); agentic-era positioning. capabilities: [tools, multimodal, agentic, long_context] placement_confidence: high notable: > Agentic-era flagship. Native image-out replaces dedicated image gen models. source_refs: [src-gemini-2-blog] - id: gemma-3 name: Gemma 3 vendor: Google release_date: "2025-03-12" year_bucket: 2025 branch: multimodal secondary_tags: [open_weights, edge, multilingual, long_context] parents: [gemma-2] influences: [gemini-2] openness: open_weights parameters: "1B / 4B / 12B / 27B" context: "128K" modalities: [text, image] architecture_notes: > Built on Gemini 2.0 tech; multimodal (4B/12B/27B); 140+ languages; function calling. capabilities: [tools, multimodal, open_weights, long_context] placement_confidence: high notable: > Apache-2.0 licensed; fits on a single GPU/TPU. source_refs: [src-gemma-3-blog] - id: gemma-4 name: Gemma 4 vendor: Google release_date: "2026-03-31" year_bucket: 2026 branch: multimodal secondary_tags: [open_weights] parents: [gemma-3] influences: [gemini-3-pro] openness: open_weights parameters: "Multiple sizes (specifics not yet documented)" context: "Not yet documented" modalities: [text, image] architecture_notes: > Successor to Gemma 3; leverages Gemini 3 research generation. capabilities: [open_weights, multimodal] placement_confidence: medium notable: > Latest Gemma sibling line; details still being published. source_refs: [src-gemma-releases] - id: llama-3-2-vision name: Llama 3.2 Vision (11B / 90B) vendor: Meta AI release_date: "2024-09-25" year_bucket: 2024 branch: multimodal secondary_tags: [open_weights, long_context] parents: [llama-3-1] influences: [] openness: open_weights parameters: "11B, 90B" context: "128K" modalities: [text, image] architecture_notes: > Vision adapter with cross-attention atop the Llama 3.1 text model; trained on 6B image-text pairs. capabilities: [multimodal, open_weights, long_context] placement_confidence: high notable: > Meta's first openly-released vision-language Llamas. source_refs: [src-llama-3-2-blog] - id: pixtral-12b name: Pixtral 12B vendor: Mistral AI release_date: "2024-09-17" year_bucket: 2024 branch: multimodal secondary_tags: [open_weights, eu, long_context] parents: [mistral-nemo-12b] influences: [llama-3-2-vision] openness: open_weights parameters: "12B + 400M vision encoder" context: "128K" modalities: [text, image] architecture_notes: > Apache 2.0; variable image sizes/aspect ratios. capabilities: [multimodal, open_weights, long_context] placement_confidence: high notable: > Mistral's first multimodal model. source_refs: [src-pixtral-blog] - id: pixtral-large name: Pixtral Large (124B) vendor: Mistral AI release_date: "2024-11-18" year_bucket: 2024 branch: multimodal secondary_tags: [open_weights, eu, long_context] parents: [mistral-large-2] influences: [pixtral-12b] openness: open_weights parameters: "124B" context: "128K" modalities: [text, image] architecture_notes: > MRL license; built on Mistral Large 2. capabilities: [multimodal, open_weights, long_context] placement_confidence: high notable: > Frontier open-weights vision model. source_refs: [src-pixtral-large-blog] - id: voxtral name: Voxtral 24B / 3B vendor: Mistral AI release_date: "2025-07-15" year_bucket: 2025 branch: multimodal secondary_tags: [open_weights, audio, eu] parents: [mistral-small-3] influences: [] openness: open_weights parameters: "24B, 3B" context: "32K (up to 30-40 min audio)" modalities: [text, audio] architecture_notes: > Apache 2.0; native multilingual ASR plus understanding; function-calling from voice. capabilities: [multimodal, open_weights] placement_confidence: high notable: > Western open-weights speech-understanding frontier. source_refs: [src-voxtral-blog] - id: qwen-2-5-vl name: Qwen2.5-VL vendor: Alibaba release_date: "2025-01" year_bucket: 2025 branch: multimodal secondary_tags: [open_weights, video, china, long_context] parents: [qwen-2-5] influences: [] openness: open_weights parameters: "3B / 7B / 72B" context: "128K" modalities: [text, image, video] architecture_notes: > Qwen2.5 family with vision plus video. capabilities: [multimodal, open_weights, long_context] placement_confidence: high notable: > Strong open-weights vision-language baseline for the Chinese frontier. source_refs: [src-qwen-blog] - id: qwen-3-5 name: Qwen3.5 vendor: Alibaba release_date: "2026-02-24" year_bucket: 2026 branch: multimodal secondary_tags: [open_weights, mixture_of_experts, agentic, china, long_context] parents: [qwen-3] influences: [] openness: open_weights parameters: "Multiple sizes incl. ~397B reasoning variant" context: "128K" modalities: [text, image, video] architecture_notes: > Unified vision-language foundation, early-fusion training; Gated Delta Networks plus sparse MoE; RL across million-agent environments; 201 languages. capabilities: [tools, multimodal, agentic, open_weights, long_context] placement_confidence: high notable: > Bridge from Qwen3 to the 3.6 family; multilingual + multimodal scale. source_refs: [src-qwen-3-6-repo] - id: qwen-3-6-plus name: Qwen3.6-Plus vendor: Alibaba release_date: "2026-04-02" year_bucket: 2026 branch: multimodal secondary_tags: [closed, agentic, frontier, china, ultra_long_context] parents: [qwen-3-5] influences: [] openness: closed parameters: "Undisclosed (flagship Plus)" context: "1M" modalities: [text, image, video] architecture_notes: > Visual coding (screenshots to code); repository-level engineering; document/visual/video reasoning. capabilities: [tools, multimodal, agentic, long_context] placement_confidence: high notable: > Alibaba's flagship agentic multimodal model. source_refs: [src-qwen-3-6-plus-press] - id: qwen-3-6-35b-a3b name: Qwen3.6-35B-A3B vendor: Alibaba release_date: "2026-04-16" year_bucket: 2026 branch: multimodal secondary_tags: [open_weights, mixture_of_experts, agentic, china, long_context] parents: [qwen-3-6-plus] influences: [] openness: open_weights parameters: "35B / 3B active" context: "128K" modalities: [text, image] architecture_notes: > Open-weights MoE sibling of Qwen3.6-Plus. capabilities: [tools, multimodal, agentic, open_weights, long_context] placement_confidence: high notable: > 73.4% SWE-bench Verified at 3B active parameters. source_refs: [src-qwen-3-6-35b-news] - id: doubao-1-5 name: Doubao 1.5 (Deep Thinking) vendor: ByteDance release_date: "2025-04" year_bucket: 2025 branch: multimodal secondary_tags: [closed, mixture_of_experts, video, china] parents: [] influences: [deepseek-r1] openness: closed parameters: "20B total / 2B active" context: "Undisclosed" modalities: [text, image, video] architecture_notes: > ByteDance's first multimodal MoE with thinking mode. capabilities: [multimodal] placement_confidence: high notable: > Anchor of the Doubao multimodal line. source_refs: [src-doubao-1-5-news] - id: doubao-1-6-vision name: Doubao 1.6-Vision vendor: ByteDance release_date: "2025-09" year_bucket: 2025 branch: multimodal secondary_tags: [closed, tool_using, china] parents: [doubao-1-5] influences: [] openness: closed parameters: "Undisclosed" context: "Undisclosed" modalities: [text, image] architecture_notes: > Vision-focused refinement with tool use. capabilities: [tools, multimodal] placement_confidence: high notable: > Bridge to the variable-thinking Doubao 1.6. source_refs: [src-doubao-1-6-vision-news] - id: ernie-4-5 name: ERNIE 4.5 vendor: Baidu release_date: "2025-03-16" year_bucket: 2025 branch: multimodal secondary_tags: [closed, china] parents: [ernie-4] influences: [] openness: closed parameters: "Undisclosed" context: "128K" modalities: [text, image] architecture_notes: > Baidu's first native multimodal foundation model. capabilities: [multimodal, long_context] placement_confidence: high notable: > Anchor of the ERNIE multimodal line. source_refs: [src-ernie-4-5-news] - id: ernie-5 name: ERNIE 5.0 vendor: Baidu release_date: "2026-01-22" year_bucket: 2026 branch: multimodal secondary_tags: [closed, frontier, china] parents: [ernie-4-5] influences: [] openness: closed parameters: "2.4T (claimed)" context: "Undisclosed" modalities: [text, image, audio, video] architecture_notes: > Native full-modal I/O — text plus image plus audio plus video in/out at frontier scale. capabilities: [multimodal] placement_confidence: medium notable: > Preview Nov 2025; GA Jan 22, 2026; AIME 2025 87.0%. source_refs: [src-ernie-5-baike] - id: minimax-vl-01 name: MiniMax-VL-01 vendor: MiniMax release_date: "2025-01-15" year_bucket: 2025 branch: multimodal secondary_tags: [open_weights, mixture_of_experts, china, ultra_long_context] parents: [minimax-text-01] influences: [] openness: open_weights parameters: "456B total / 45.9B active" context: "1M" modalities: [text, image] architecture_notes: > Lightning Attention multimodal sibling of MiniMax-Text-01. capabilities: [multimodal, open_weights, long_context] placement_confidence: high notable: > Carried Lightning Attention to multimodal at frontier scale. source_refs: [src-minimax-vl-repo] - id: phi-4-multimodal name: Phi-4-multimodal / Phi-4-mini vendor: Microsoft release_date: "2025-02-26" year_bucket: 2025 branch: multimodal secondary_tags: [open_weights, small, distilled, audio, long_context] parents: [phi-4] influences: [] openness: open_weights parameters: "5.6B (multimodal); 3.8B (mini)" context: "128K" modalities: [text, image, audio] architecture_notes: > Mixture-of-LoRAs design; first Phi with unified speech/vision/text representation. capabilities: [multimodal, open_weights, long_context] placement_confidence: high notable: > Phi family enters the unified omni-modal small-model space. source_refs: [src-phi-4-mm-blog] - id: falcon-2-11b name: Falcon 2 11B (+ VLM) vendor: TII release_date: "2024-05-13" year_bucket: 2024 branch: multimodal secondary_tags: [open_weights, uae] parents: [falcon-180b] influences: [] openness: open_weights parameters: "11B" context: "8K" modalities: [text, image] architecture_notes: > Trained on 5T+ RefinedWeb tokens; VLM variant adds image input. capabilities: [multimodal, open_weights] placement_confidence: high notable: > TII's first multimodal release. source_refs: [src-falcon-2-blog] - id: granite-3-2 name: IBM Granite 3.2 (incl. Granite Vision 3.2 2B) vendor: IBM release_date: "2025-02-26" year_bucket: 2025 branch: multimodal secondary_tags: [open_weights, enterprise, long_context] parents: [granite-3-1] influences: [] openness: open_weights parameters: "8B/2B Instruct; Vision 2B" context: "128K" modalities: [text, image] architecture_notes: > Toggleable chain-of-thought reasoning; Vision 2B trained on 85M PDFs plus 26M synthetic QA. capabilities: [multimodal, open_weights, long_context] placement_confidence: high notable: > First Granite multimodal; document-understanding focused vision model. source_refs: [src-granite-3-2-blog] # ============================================================ # REASONING # ============================================================ - id: o1-preview name: OpenAI o1-preview vendor: OpenAI release_date: "2024-09-12" year_bucket: 2024 branch: reasoning secondary_tags: [closed, frontier, historical_frontier] parents: [gpt-4o] influences: [gpt-4-turbo] openness: closed parameters: "undisclosed" context: "128K" modalities: [text] architecture_notes: > First public OpenAI model trained with large-scale RL to produce long internal chain-of-thought before answering. capabilities: [reasoning] placement_confidence: high notable: > Launched the o-series. 83% on IMO qualifying exams vs. 13% for GPT-4o. source_refs: [src-o1-preview-blog] - id: o1 name: OpenAI o1 vendor: OpenAI release_date: "2024-12-05" year_bucket: 2024 branch: reasoning secondary_tags: [closed, historical_frontier, multimodal] parents: [o1-preview] influences: [] openness: closed parameters: "undisclosed" context: "128K-200K" modalities: [text, image] architecture_notes: > GA version of the reasoning model. Added vision support and function calling, structured outputs. capabilities: [tools, multimodal, reasoning] placement_confidence: high notable: > First reasoning model with full API support. source_refs: [src-o1-blog] - id: o1-mini name: OpenAI o1-mini vendor: OpenAI release_date: "2024-09-12" year_bucket: 2024 branch: reasoning secondary_tags: [closed, small] parents: [o1-preview] influences: [] openness: closed parameters: "undisclosed (small)" context: "128K" modalities: [text] architecture_notes: > Smaller reasoning model optimized for STEM/code. capabilities: [reasoning] placement_confidence: high notable: > Made reasoning models accessible at low cost. source_refs: [src-o1-mini-blog] - id: o1-pro name: OpenAI o1-pro vendor: OpenAI release_date: "2024-12-05" year_bucket: 2024 branch: reasoning secondary_tags: [closed, multimodal] parents: [o1] influences: [] openness: closed parameters: "undisclosed" context: "128K-200K" modalities: [text, image] architecture_notes: > Same underlying model as o1, configured to spend longer at inference. capabilities: [reasoning, multimodal] placement_confidence: medium notable: > Introduced the pro (extended-thinking) variant pattern. source_refs: [src-o1-pro-blog] - id: o3-mini name: OpenAI o3-mini vendor: OpenAI release_date: "2025-01-31" year_bucket: 2025 branch: reasoning secondary_tags: [closed, small, tool_using] parents: [o1-mini] influences: [o1] openness: closed parameters: "undisclosed (small)" context: "200K" modalities: [text] architecture_notes: > Three configurable reasoning-effort levels (low/medium/high). Function calling, structured outputs, streaming. capabilities: [tools, reasoning] placement_confidence: high notable: > First reasoning model offered to ChatGPT Free users. source_refs: [src-o3-mini-blog] - id: o3 name: OpenAI o3 vendor: OpenAI release_date: "2025-04-16" year_bucket: 2025 branch: reasoning secondary_tags: [closed, frontier, agentic, multimodal, tool_using] parents: [o1] influences: [o3-mini] openness: closed parameters: "undisclosed" context: "200K" modalities: [text, image] architecture_notes: > First reasoning model that could autonomously chain tool calls inside its chain-of-thought. SOTA on Codeforces, SWE-bench, MMMU at release. capabilities: [tools, multimodal, reasoning, agentic] placement_confidence: high notable: > 20% fewer major errors than o1. source_refs: [src-o3-blog] - id: o4-mini name: OpenAI o4-mini vendor: OpenAI release_date: "2025-04-16" year_bucket: 2025 branch: reasoning secondary_tags: [closed, small, agentic, multimodal, tool_using] parents: [o3-mini] influences: [o3] openness: closed parameters: "undisclosed (small)" context: "200K" modalities: [text, image] architecture_notes: > Small fast reasoning model with tool use. Note: there is no o4 (full) — only o4-mini exists. capabilities: [tools, multimodal, reasoning, agentic] placement_confidence: high notable: > 99.5% pass@1 on AIME 2025 with tools. source_refs: [src-o3-blog] - id: o3-pro name: OpenAI o3-pro vendor: OpenAI release_date: "2025-06-10" year_bucket: 2025 branch: reasoning secondary_tags: [closed, agentic, multimodal, tool_using] parents: [o3] influences: [o1-pro] openness: closed parameters: "undisclosed" context: "200K" modalities: [text, image] architecture_notes: > Same underlying o3 with extended thinking time. Replaces o1-pro. Image generation NOT supported. capabilities: [tools, multimodal, reasoning, agentic] placement_confidence: high notable: > Last pro variant before the GPT-5 line absorbed the extended-thinking pattern. source_refs: [src-o3-pro-blog] - id: gpt-5 name: GPT-5 vendor: OpenAI release_date: "2025-08-07" year_bucket: 2025 branch: reasoning secondary_tags: [closed, frontier, agentic, multimodal, mcp, tool_using, ultra_long_context] parents: [gpt-4-5, o3] influences: [gpt-4o, o4-mini] openness: closed parameters: "undisclosed" context: "400K (128K max output)" modalities: [text, image, audio] architecture_notes: > Unified system: GPT-5 Main (fast) plus GPT-5 Thinking (deep reasoning) plus a real-time router. NOT mixture-of-experts. capabilities: [tools, mcp, multimodal, reasoning, agentic, long_context] placement_confidence: high notable: > Knit together OpenAI's two parallel lineages (pretraining-scaling and reasoning) into one product. source_refs: [src-gpt-5-blog] - id: gpt-5-codex name: GPT-5-Codex vendor: OpenAI release_date: "2025-09-15" year_bucket: 2025 branch: reasoning secondary_tags: [closed, code_specialist, agentic, mcp, tool_using] parents: [gpt-5] influences: [] openness: closed parameters: "undisclosed" context: "400K" modalities: [text, image] architecture_notes: > GPT-5 fine-tuned for agentic coding. capabilities: [tools, mcp, multimodal, reasoning, agentic, long_context] placement_confidence: high notable: > Reuses the Codex brand from the 2021 GPT-3 code model but is a fundamentally different lineage. source_refs: [src-gpt-5-codex-blog] - id: gpt-5-1 name: GPT-5.1 (Instant + Thinking + Pro) vendor: OpenAI release_date: "2025-11-13" year_bucket: 2025 branch: reasoning secondary_tags: [closed, frontier, agentic, multimodal, ultra_long_context] parents: [gpt-5] influences: [] openness: closed parameters: "undisclosed" context: "400K" modalities: [text, image, audio] architecture_notes: > First post-GPT-5 minor refresh. Instant and Thinking variants separated as user-selectable. capabilities: [tools, mcp, multimodal, reasoning, agentic, long_context] placement_confidence: high notable: > Established the Instant vs. Thinking UI split that all subsequent 5.x models inherit. source_refs: [src-gpt-5-1-blog] - id: gpt-5-2 name: GPT-5.2 vendor: OpenAI release_date: "2025-12-11" year_bucket: 2025 branch: reasoning secondary_tags: [closed, frontier, multimodal, mcp, ultra_long_context] parents: [gpt-5-1] influences: [] openness: closed parameters: "undisclosed" context: "400K" modalities: [text, image, audio] architecture_notes: > Higher performance across reasoning, coding, long context, vision. Powers ChatGPT Deep Research from Feb 11, 2026. capabilities: [tools, mcp, multimodal, reasoning, agentic, long_context] placement_confidence: high notable: > Last 5.x model to remain alongside 5.3 — Instant tier now uses 5.3, Thinking tier moved to 5.4. source_refs: [src-gpt-5-2-blog] - id: gpt-5-3-instant name: GPT-5.3 Instant vendor: OpenAI release_date: "2026-03-03" year_bucket: 2026 branch: reasoning secondary_tags: [closed, multimodal, mcp, ultra_long_context] parents: [gpt-5-2] influences: [] openness: closed parameters: "undisclosed" context: "400K" modalities: [text, image, audio] architecture_notes: > Successor to GPT-5.2 Instant. Focused on tone, relevance, conversational flow rather than benchmark gains. capabilities: [tools, mcp, multimodal, reasoning, agentic, long_context] placement_confidence: high notable: > OpenAI did NOT release a GPT-5.3 Thinking — the next thinking-tier release was GPT-5.4 Thinking. source_refs: [src-gpt-5-3-instant-blog] - id: gpt-5-4 name: GPT-5.4 (Thinking + Pro + mini + nano) vendor: OpenAI release_date: "2026-03-05" year_bucket: 2026 branch: reasoning secondary_tags: [closed, frontier, agentic, multimodal, code_specialist, mcp, ultra_long_context] parents: [gpt-5-2, gpt-5-codex] influences: [gpt-5-3-instant] openness: closed parameters: "undisclosed" context: "400K+" modalities: [text, image, audio] architecture_notes: > Unified frontier model integrating reasoning plus GPT-5.3-Codex coding capabilities plus agentic workflows. capabilities: [tools, mcp, multimodal, reasoning, agentic, long_context] placement_confidence: high notable: > First explicit fusion of the GPT-5 base plus Codex training stacks into the main flagship. source_refs: [src-gpt-5-4-blog] - id: gpt-5-5 name: GPT-5.5 vendor: OpenAI release_date: "2026-04-23" year_bucket: 2026 branch: reasoning secondary_tags: [closed, agentic, frontier, multimodal, ultra_long_context] parents: [gpt-5-4] influences: [] openness: closed parameters: "undisclosed" context: "400K+" modalities: [text, image, audio] architecture_notes: > Positioned as a new class of intelligence for real work — autonomous multi-part task completion. Same per-token latency as GPT-5.4 but uses meaningfully fewer tokens to complete tasks. capabilities: [tools, mcp, multimodal, reasoning, agentic, long_context] placement_confidence: high notable: > Released two days before this dataset (April 23, 2026). Terminal-Bench 2.0: 82.7%. GDPval (44 occupations): 84.9%. OSWorld-Verified: 78.7%. source_refs: [src-gpt-5-5-blog] # ---- Anthropic reasoning era ---- - id: claude-3-7-sonnet name: Claude 3.7 Sonnet vendor: Anthropic release_date: "2025-02-24" year_bucket: 2025 branch: reasoning secondary_tags: [closed, multimodal, hybrid_reasoning, agentic, tool_using, long_context] parents: [claude-3-5-sonnet] influences: [o1] openness: closed parameters: "undisclosed" context: "200K" modalities: [text, image] architecture_notes: > Dense decoder; first hybrid reasoning Claude — single model that can produce instant responses or show extended thinking step-by-step. capabilities: [tools, multimodal, reasoning, long_context] placement_confidence: high notable: > Anthropic's first hybrid reasoning model. source_refs: [src-claude-3-7-sonnet-blog] - id: claude-opus-4 name: Claude Opus 4 vendor: Anthropic release_date: "2025-05-22" year_bucket: 2025 branch: reasoning secondary_tags: [closed, frontier, multimodal, extended_thinking, agentic, mcp, tool_using, long_context] parents: [claude-3-7-sonnet] influences: [] openness: closed parameters: "undisclosed" context: "200K" modalities: [text, image] architecture_notes: > Dense decoder; extended-thinking by default. Interleaved thinking (think between tool calls) added as beta. capabilities: [tools, mcp, multimodal, reasoning, agentic, long_context] placement_confidence: high notable: > Top-tier coding model for long-running agentic tasks. source_refs: [src-claude-4-blog] - id: claude-sonnet-4 name: Claude Sonnet 4 vendor: Anthropic release_date: "2025-05-22" year_bucket: 2025 branch: reasoning secondary_tags: [closed, multimodal, extended_thinking, mcp, tool_using, ultra_long_context] parents: [claude-3-7-sonnet] influences: [] openness: closed parameters: "undisclosed" context: "200K (1M beta added Aug 12, 2025)" modalities: [text, image] architecture_notes: > Extended thinking, interleaved thinking, code execution, MCP connector, Files API. capabilities: [tools, mcp, multimodal, reasoning, long_context] placement_confidence: high notable: > First Claude to ship with 1M-token context (beta, Aug 2025). source_refs: [src-claude-4-blog] - id: claude-opus-4-1 name: Claude Opus 4.1 vendor: Anthropic release_date: "2025-08-05" year_bucket: 2025 branch: reasoning secondary_tags: [closed, frontier, multimodal, extended_thinking, agentic, mcp, tool_using, long_context] parents: [claude-opus-4] influences: [] openness: closed parameters: "undisclosed" context: "200K" modalities: [text, image] architecture_notes: > Dense decoder; incremental upgrade over Opus 4. capabilities: [tools, mcp, multimodal, reasoning, agentic, long_context] placement_confidence: high notable: > Improved coding and reasoning over Opus 4. source_refs: [src-claude-opus-4-1-blog] - id: claude-sonnet-4-5 name: Claude Sonnet 4.5 vendor: Anthropic release_date: "2025-09-29" year_bucket: 2025 branch: reasoning secondary_tags: [closed, multimodal, agentic, extended_thinking, computer_use, mcp, tool_using, long_context] parents: [claude-sonnet-4] influences: [] openness: closed parameters: "undisclosed" context: "200K" modalities: [text, image] architecture_notes: > Positioned as best model for complex agents and coding. Memory tool and context editing introduced alongside. capabilities: [tools, mcp, multimodal, reasoning, agentic, computer_use, long_context] placement_confidence: high notable: > Anthropic's flagship for agents/coding/computer-use through late 2025. source_refs: [src-claude-sonnet-4-5-blog] - id: claude-haiku-4-5 name: Claude Haiku 4.5 vendor: Anthropic release_date: "2025-10-15" year_bucket: 2025 branch: reasoning secondary_tags: [closed, small, multimodal, extended_thinking, mcp, tool_using, long_context] parents: [claude-3-5-haiku] influences: [claude-sonnet-4] openness: closed parameters: "undisclosed" context: "200K (64K output)" modalities: [text, image] architecture_notes: > Near-frontier performance at Haiku-tier price. Sonnet-4-class coding and agentic capability at ~1/3 the cost and 2x the speed. capabilities: [tools, mcp, multimodal, reasoning, agentic, long_context] placement_confidence: high notable: > First Haiku with extended-thinking. Skipped a Haiku 4 SKU. source_refs: [src-claude-haiku-4-5-blog] - id: claude-opus-4-5 name: Claude Opus 4.5 vendor: Anthropic release_date: "2025-11-24" year_bucket: 2025 branch: reasoning secondary_tags: [closed, frontier, multimodal, extended_thinking, agentic, computer_use, mcp, tool_using, long_context] parents: [claude-opus-4-1] influences: [] openness: closed parameters: "undisclosed" context: "200K (1M beta added Feb 5, 2026)" modalities: [text, image] architecture_notes: > Introduced effort parameter (low/medium/high/max) for thinking control, programmatic tool calling, tool search. capabilities: [tools, mcp, multimodal, reasoning, agentic, computer_use, long_context] placement_confidence: high notable: > 80.9% on SWE-bench Verified at launch. Pricing dropped to $5/$25 per MTok — a third of prior Opus pricing. source_refs: [src-claude-opus-4-5-blog] - id: claude-opus-4-6 name: Claude Opus 4.6 vendor: Anthropic release_date: "2026-02-05" year_bucket: 2026 branch: reasoning secondary_tags: [closed, frontier, multimodal, adaptive_thinking, agentic, computer_use, mcp, tool_using, ultra_long_context] parents: [claude-opus-4-5] influences: [] openness: closed parameters: "undisclosed" context: "1M (GA March 13, 2026)" modalities: [text, image] architecture_notes: > Introduced ADAPTIVE THINKING — the model auto-decides reasoning depth. Compaction API, fast mode, agent teams. capabilities: [tools, mcp, multimodal, reasoning, agentic, computer_use, long_context] placement_confidence: high notable: > First Opus with 1M-token context. 76% MRCR-v2 retrieval at 1M tokens. source_refs: [src-claude-opus-4-6-blog] - id: claude-sonnet-4-6 name: Claude Sonnet 4.6 vendor: Anthropic release_date: "2026-02-17" year_bucket: 2026 branch: reasoning secondary_tags: [closed, multimodal, extended_thinking, agentic, computer_use, mcp, tool_using, ultra_long_context] parents: [claude-sonnet-4-5] influences: [claude-opus-4-6] openness: closed parameters: "undisclosed" context: "1M (GA March 13, 2026)" modalities: [text, image] architecture_notes: > Full upgrade across coding, computer use, long-context reasoning, agent planning, design. Better agentic search at lower token consumption. capabilities: [tools, mcp, multimodal, reasoning, agentic, computer_use, long_context] placement_confidence: high notable: > Sonnet-tier 1M-context model. source_refs: [src-claude-sonnet-4-6-blog] - id: claude-mythos name: Claude Mythos Preview vendor: Anthropic release_date: "2026-04-07" year_bucket: 2026 branch: reasoning secondary_tags: [gated, frontier, multimodal, agentic] parents: [claude-opus-4-6] influences: [] openness: gated parameters: "undisclosed" context: "undisclosed" modalities: [text, image] architecture_notes: > Anthropic's most capable model as of April 2026. Architecture not publicly disclosed. Specialized capability profile for autonomous vulnerability discovery and exploit generation. capabilities: [tools, multimodal, reasoning, agentic] placement_confidence: medium notable: > Invitation-only research preview as part of Project Glasswing. 83.1% on CyberGym vs. Opus 4.6's ~74%. NOT generally available. source_refs: [src-claude-mythos-page, src-mythos-greek-news] status: gated - id: claude-opus-4-7 name: Claude Opus 4.7 vendor: Anthropic release_date: "2026-04-16" year_bucket: 2026 branch: reasoning secondary_tags: [closed, frontier, multimodal, adaptive_thinking, agentic, computer_use, mcp, tool_using, ultra_long_context] parents: [claude-opus-4-6] influences: [claude-mythos] openness: closed parameters: "undisclosed" context: "1M" modalities: [text, image] architecture_notes: > Adaptive thinking plus new xhigh effort level between high and max. Updated tokenizer (1.0–1.35x input token count vs. Opus 4.6). Vision lifted to 2,576 px long-edge / ~3.75 MP (3x prior limit). capabilities: [tools, mcp, multimodal, reasoning, agentic, computer_use, long_context] placement_confidence: high notable: > Most capable generally available Claude as of April 25, 2026 (Mythos Preview is more capable but gated). source_refs: [src-claude-opus-4-7-blog] - id: claude-opus-4-8 name: Claude Opus 4.8 vendor: Anthropic release_date: "2026-05-28" year_bucket: 2026 branch: reasoning secondary_tags: [closed, frontier, multimodal, adaptive_thinking, agentic, computer_use, mcp, tool_using, ultra_long_context] parents: [claude-opus-4-7] influences: [claude-mythos] openness: closed parameters: "undisclosed" context: "1M" modalities: [text, image] architecture_notes: > Adds selectable effort control as a first-class GA feature and Dynamic Workflows that orchestrate up to 1,000 parallel subagents for codebase-scale tasks. Cacheable-prompt minimum lowered to 1,024 tokens. capabilities: [tools, mcp, multimodal, reasoning, agentic, computer_use, long_context] placement_confidence: high notable: > Retook the #1 spot on the Artificial Analysis Intelligence Index (61.4 vs GPT-5.5's 60.2) on May 28, 2026, with SWE-Bench Pro 69.2%; headline pricing held at $5/$25 per Mtok while fast mode dropped ~3x. source_refs: [src-claude-opus-4-8-blog] # ---- Google reasoning era ---- - id: gemini-2-5-pro name: Gemini 2.5 Pro vendor: Google + DeepMind release_date: "2025-03-25" year_bucket: 2025 branch: reasoning secondary_tags: [closed, mixture_of_experts, multimodal, ultra_long_context] parents: [gemini-2] influences: [] openness: closed parameters: "Undisclosed (sparse MoE)" context: "1M" modalities: [text, image, audio, video] architecture_notes: > Sparse MoE transformer with native multimodality; trained to generate explicit reasoning traces before final response (Deep Think mode). capabilities: [tools, multimodal, reasoning, agentic, long_context] placement_confidence: high notable: > First Gemini thinking model; debuted #1 on LMArena. source_refs: [src-gemini-2-5-blog] - id: gemini-3-pro name: Gemini 3 Pro vendor: Google + DeepMind release_date: "2025-11-18" year_bucket: 2025 branch: reasoning secondary_tags: [closed, mixture_of_experts, multimodal, agentic, frontier, ultra_long_context] parents: [gemini-2-5-pro] influences: [] openness: closed parameters: "Undisclosed" context: "1M+" modalities: [text, image, audio, video] architecture_notes: > Successor sparse-MoE thinking model; integrated Deep Think mode. capabilities: [tools, multimodal, reasoning, agentic, long_context] placement_confidence: high notable: > 91.9% GPQA Diamond, 23.4% MathArena Apex, 81% MMMU-Pro. source_refs: [src-gemini-3-card] - id: gemini-3-1-pro name: Gemini 3.1 Pro vendor: Google + DeepMind release_date: "2026-02-19" year_bucket: 2026 branch: reasoning secondary_tags: [closed, multimodal, agentic, frontier, ultra_long_context] parents: [gemini-3-pro] influences: [] openness: closed parameters: "Undisclosed" context: "1M+" modalities: [text, image, audio, video] architecture_notes: > Iterative refresh on Gemini 3 Pro; powers Google's research agents and Personal Intelligence features. capabilities: [tools, multimodal, reasoning, agentic, long_context] placement_confidence: high notable: > Underpins Google's April 2026 AI research agents product. source_refs: [src-gemini-3-1-pro-news] # ---- DeepSeek reasoning era ---- - id: deepseek-r1 name: DeepSeek-R1 vendor: DeepSeek AI release_date: "2025-01-20" year_bucket: 2025 branch: reasoning secondary_tags: [open_weights, mixture_of_experts, frontier, china, long_context] parents: [deepseek-v3] influences: [o1] openness: open_weights parameters: "671B total / 37B active" context: "128K" modalities: [text] architecture_notes: > GRPO-based RL post-training; pure RL (R1-Zero) and cold-start variant (R1). capabilities: [reasoning, open_weights, long_context] placement_confidence: high notable: > Open-weights reasoning model that triggered the 2025 efficiency-shock cycle. source_refs: [src-deepseek-r1-paper, src-deepseek-r1-nature] - id: deepseek-r1-distill name: DeepSeek-R1-Distill (Qwen / Llama family) vendor: DeepSeek AI release_date: "2025-01-22" year_bucket: 2025 branch: reasoning secondary_tags: [open_weights, distilled, china] parents: [deepseek-r1] influences: [qwen-2-5, llama-3] openness: open_weights parameters: "1.5B / 7B / 8B / 14B / 32B / 70B" context: "32K-128K" modalities: [text] architecture_notes: > SFT on 800K CoT samples drawn from R1 onto Qwen2.5 and Llama-3.x backbones. capabilities: [reasoning, open_weights] placement_confidence: high notable: > Made R1-class reasoning available in small open-weights sizes. source_refs: [src-deepseek-r1-distill-card] - id: deepseek-r2 name: DeepSeek-R2 vendor: DeepSeek AI release_date: "2026-04" year_bucket: 2026 branch: reasoning secondary_tags: [open_weights, frontier, china, long_context] parents: [deepseek-r1] influences: [deepseek-v3-2] openness: open_weights parameters: "32B (dense)" context: "128K" modalities: [text] architecture_notes: > Pivot from rumored 1.2T MoE to 32B dense after Huawei Ascend training failures forced reversion to NVIDIA GPUs. capabilities: [reasoning, open_weights, long_context] placement_confidence: medium notable: > Strategic shift to consumer-GPU-runnable reasoning; 92.7% AIME 2025. source_refs: [src-deepseek-r2-page] # ---- Qwen / Alibaba reasoning ---- - id: qwq-32b-preview name: QwQ-32B-Preview vendor: Alibaba release_date: "2024-11" year_bucket: 2024 branch: reasoning secondary_tags: [open_weights, china] parents: [qwen-2-5] influences: [o1] openness: open_weights parameters: "32B" context: "32K" modalities: [text] architecture_notes: > Apache 2.0; first Alibaba o1-style preview. capabilities: [reasoning, open_weights] placement_confidence: high notable: > Alibaba's first reasoning-model entry. source_refs: [src-qwen-3-blog] - id: qwen-3-thinking-2507 name: Qwen3-235B-A22B-Thinking-2507 vendor: Alibaba release_date: "2025-07" year_bucket: 2025 branch: reasoning secondary_tags: [open_weights, mixture_of_experts, china, long_context] parents: [qwen-3] influences: [] openness: open_weights parameters: "235B / 22B active" context: "128K" modalities: [text] architecture_notes: > Reasoning-mode variant of Qwen3 flagship MoE. capabilities: [reasoning, open_weights, long_context] placement_confidence: high notable: > Qwen3 reasoning-mode flagship. source_refs: [src-qwen-3-paper] # ---- Kimi reasoning ---- - id: kimi-k1-5 name: Kimi K1.5 vendor: Moonshot AI release_date: "2025-01-20" year_bucket: 2025 branch: reasoning secondary_tags: [closed, multimodal, china, long_context] parents: [kimi-k1] influences: [o1] openness: closed parameters: "Undisclosed" context: "128K" modalities: [text, image] architecture_notes: > RL-trained CoT; o1-level multimodal; released same week as DeepSeek-R1. capabilities: [reasoning, multimodal, long_context] placement_confidence: high notable: > First Kimi reasoning model. source_refs: [src-kimi-k1-5-page] - id: kimi-k2-5 name: Kimi K2.5 vendor: Moonshot AI release_date: "2026-01-26" year_bucket: 2026 branch: reasoning secondary_tags: [open_weights, mixture_of_experts, multimodal, agentic, china, long_context] parents: [kimi-k2] influences: [] openness: open_weights parameters: "1T total / 32B active" context: "256K" modalities: [text, image] architecture_notes: > Native vision (silent rollout); thinking-mode tool calling. capabilities: [tools, reasoning, multimodal, agentic, open_weights, long_context] placement_confidence: high notable: > Bridge from Kimi K2 to the K2.6 multi-agent flagship. source_refs: [src-kimi-k2-5-blog] # ---- GLM reasoning ---- - id: glm-z1 name: GLM-Z1-32B-0414 vendor: Z.AI (Zhipu) release_date: "2025-04" year_bucket: 2025 branch: reasoning secondary_tags: [open_weights, china, long_context] parents: [glm-4-32b-0414] influences: [deepseek-r1] openness: open_weights parameters: "32B (also 9B variant)" context: "128K" modalities: [text] architecture_notes: > Cold start plus extended RL; sister Rumination variant uses end-to-end RL with search-tool calls. capabilities: [tools, reasoning, open_weights, long_context] placement_confidence: high notable: > Zhipu's open-weights reasoning entry. source_refs: [src-glm-z1-docs] - id: glm-5-1 name: GLM-5.1 vendor: Z.AI (Zhipu) release_date: "2026-04-08" year_bucket: 2026 branch: reasoning secondary_tags: [open_weights, mixture_of_experts, agentic, multi_agent_native, frontier, china, long_context] parents: [glm-5] influences: [] openness: open_weights parameters: "754B total" context: "203K" modalities: [text] architecture_notes: > Asynchronous RL infrastructure decoupling generation from training; sustains 8-hour autonomous execution. capabilities: [tools, reasoning, agentic, multi_agent_native, open_weights, long_context] placement_confidence: high notable: > SOTA SWE-Bench Pro 58.4 — beats GPT-5.4, Claude Opus 4.6, Gemini 3.1 Pro. source_refs: [src-glm-5-1-news] # ---- Doubao / ByteDance reasoning ---- - id: doubao-1-6 name: Doubao 1.6 vendor: ByteDance release_date: "2025-10" year_bucket: 2025 branch: reasoning secondary_tags: [closed, mixture_of_experts, multimodal, adaptive_thinking, china] parents: [doubao-1-6-vision] influences: [] openness: closed parameters: "Undisclosed" context: "Undisclosed" modalities: [text, image] architecture_notes: > First Chinese model with adjustable thinking depth (Minimal/Low/Medium/High). capabilities: [reasoning, multimodal] placement_confidence: high notable: > First Chinese adaptive-thinking model. source_refs: [src-doubao-1-6-news] # ---- Tencent Hunyuan reasoning ---- - id: hunyuan-t1 name: Hunyuan T1 vendor: Tencent release_date: "2025-03-21" year_bucket: 2025 branch: reasoning secondary_tags: [closed, ssm_hybrid, mamba, china, long_context] parents: [hunyuan-turbos] influences: [deepseek-r1] openness: closed parameters: "Undisclosed (MoE)" context: "128K" modalities: [text] architecture_notes: > 96.7% of compute in RL; first ultra-large Mamba-based reasoning model. capabilities: [reasoning, long_context] placement_confidence: high notable: > First ultra-large SSM-hybrid reasoning model. source_refs: [src-hunyuan-t1-page] # ---- MiniMax reasoning ---- - id: minimax-m1 name: MiniMax-M1 vendor: MiniMax release_date: "2025-06-16" year_bucket: 2025 branch: reasoning secondary_tags: [open_weights, mixture_of_experts, china, ultra_long_context] parents: [minimax-text-01] influences: [deepseek-r1] openness: open_weights parameters: "456B total / 45.9B active" context: "1M (80K reasoning output)" modalities: [text] architecture_notes: > World's first open-source large-scale hybrid-attention reasoning model. capabilities: [reasoning, open_weights, long_context] placement_confidence: high notable: > First open-weights Lightning-Attention reasoning model. source_refs: [src-minimax-m1-news] # ---- ERNIE reasoning ---- - id: ernie-x1 name: ERNIE X1 vendor: Baidu release_date: "2025-03-16" year_bucket: 2025 branch: reasoning secondary_tags: [closed, multimodal, china] parents: [ernie-4-5] influences: [deepseek-r1] openness: closed parameters: "Undisclosed" context: "128K" modalities: [text, image] architecture_notes: > Long reasoning chains, tool use; Baidu marketed as DeepSeek-R1 peer. capabilities: [tools, reasoning, multimodal] placement_confidence: high notable: > Baidu's reasoning-model entry. source_refs: [src-ernie-4-5-news] # ---- xAI Grok reasoning ---- - id: grok-3 name: Grok-3 (incl. Think / Big Brain) vendor: xAI release_date: "2025-02-18" year_bucket: 2025 branch: reasoning secondary_tags: [closed, frontier, ultra_long_context] parents: [grok-2] influences: [o1] openness: closed parameters: "Undisclosed" context: "1M" modalities: [text] architecture_notes: > Trained on Colossus cluster (~200K H100); Think/Big Brain modes. capabilities: [reasoning, long_context] placement_confidence: high notable: > xAI's first reasoning model. source_refs: [src-grok-history-page] - id: grok-4 name: Grok 4 vendor: xAI release_date: "2025-07-09" year_bucket: 2025 branch: reasoning secondary_tags: [closed, agentic, frontier, tool_using, long_context] parents: [grok-3] influences: [] openness: closed parameters: "Undisclosed" context: "256K" modalities: [text] architecture_notes: > RL training at scale on Colossus; native code interpreter plus web browsing; 6x compute efficiency vs Grok-3. capabilities: [tools, reasoning, agentic, long_context] placement_confidence: high notable: > Most agentic Grok at release. source_refs: [src-grok-4-news] - id: grok-4-20 name: Grok 4.20 vendor: xAI release_date: "2026-02-17" year_bucket: 2026 branch: reasoning secondary_tags: [closed, agentic, multi_agent_native, frontier, ultra_long_context] parents: [grok-4] influences: [] openness: closed parameters: "Undisclosed" context: "2M" modalities: [text] architecture_notes: > Native multi-agent mode (up to 16 sub-agents); toggleable reasoning across three API variants. capabilities: [tools, reasoning, agentic, multi_agent_native, long_context] placement_confidence: high notable: > 8x context vs Grok-4; Chatbot Arena ELO ~1493. source_refs: [src-grok-4-20-page] # ---- Mistral / Magistral reasoning ---- - id: magistral name: Magistral vendor: Mistral AI release_date: "2025-06-10" year_bucket: 2025 branch: reasoning secondary_tags: [open_weights, multilingual, eu] parents: [mistral-small-3] influences: [deepseek-r1, o1] openness: open_weights parameters: "Small 24B (open); Medium (closed preview)" context: "40K" modalities: [text] architecture_notes: > RLVR pipeline without distillation; native multilingual chain-of-thought. capabilities: [reasoning, open_weights] placement_confidence: high notable: > First fully Western open-weights reasoning model after DeepSeek-R1. source_refs: [src-magistral-blog] # ---- Microsoft Phi reasoning ---- - id: phi-4-reasoning name: Phi-4-reasoning / mini-reasoning vendor: Microsoft release_date: "2025-04-30" year_bucket: 2025 branch: reasoning secondary_tags: [open_weights, small, distilled, long_context] parents: [phi-4] influences: [o1, deepseek-r1] openness: open_weights parameters: "14B (reasoning/+); 3.8B (mini-reasoning); 15B (reasoning-vision)" context: "32K-128K" modalities: [text, image] architecture_notes: > SFT on chain-of-thought plus RL on outcomes; vision variant adds multimodal reasoning. Phi-4-Reasoning-Vision-15B (March 2026) extended this line with selective thinking. capabilities: [reasoning, open_weights, long_context] placement_confidence: high notable: > Beat OpenAI o1-mini and DeepSeek-R1-Distill-Llama-70B at 14B. source_refs: [src-phi-4-reasoning-card, src-phi-4-reasoning-vision-news] # ---- W18 (May 1, 2026) additions ---- # ---- Xiaomi MiMo (open-frontier MoE, ties Kimi K2.6 at #1 open-weights AA Index) ---- - id: mimo-v2-5-pro name: MiMo V2.5 Pro vendor: Xiaomi release_date: "2026-04-27" year_bucket: 2026 branch: mixture_of_experts secondary_tags: [open_weights, mixture_of_experts, agentic, frontier, china, ultra_long_context] parents: [] influences: [kimi-k2-6, deepseek-v4-pro] openness: open_weights parameters: "1.02T total / 42B active" context: "1M" modalities: [text] architecture_notes: > Hybrid attention combining Sliding Window Attention and Global Attention at 6:1 ratio across 70 layers (1 dense + 69 MoE, hidden size 6144). Three lightweight Multi-Token Prediction heads enable ~3x output speed; native FP8 E4M3 mixed precision; ~7x KV-cache reduction with 128-token window. capabilities: [tools, reasoning, agentic, open_weights, long_context] placement_confidence: high notable: > Ties Kimi K2.6 at #1 open-weights on Artificial Analysis Intelligence Index (54); Xiaomi shipped the 100T-free-token Orbit Plan alongside the weights to anchor evaluation. source_refs: [src-mimo-v2-5-pro] # ---- NVIDIA Nemotron (open multimodal MoE, hybrid Mamba-Transformer) ---- - id: nemotron-3-nano-omni name: Nemotron 3 Nano Omni vendor: NVIDIA release_date: "2026-04-28" year_bucket: 2026 branch: multimodal secondary_tags: [open_weights, mixture_of_experts, multimodal, edge, agentic, ssm_hybrid] parents: [] influences: [] openness: open_weights parameters: "30B total / 3B active" context: "256K" modalities: [text, image, video, audio] architecture_notes: > Hybrid Mamba-Transformer MoE backbone fused with C-RADIOv4-H vision encoder and Parakeet speech encoder for unified video/audio/image/text reasoning in a single forward path. Shipped in BF16, FP8, and NVFP4 formats at a 10:1 total/active ratio. capabilities: [multimodal, reasoning, agentic, open_weights, long_context] placement_confidence: high notable: > First-party NVIDIA open-weights multimodal MoE — the only Western entry in the open-frontier multimodal slot this week. source_refs: [src-nemotron-3-nano-omni] # ---- OpenAI Privacy Filter (first OpenAI open-weights since GPT-2) ---- - id: openai-privacy-filter name: OpenAI Privacy Filter vendor: OpenAI release_date: "2026-04-28" year_bucket: 2026 branch: encoder_only secondary_tags: [open_weights, mixture_of_experts, small, edge] parents: [] influences: [] openness: open_weights parameters: "1.5B total / 50M active" context: "128K" modalities: [text] architecture_notes: > Sparse MoE token classifier (128 experts, top-4 routing, ~30:1 total-to-active ratio). Three-phase build: GPT-style autoregressive pretraining, architectural conversion to bidirectional token classifier, supervised PII post-training. Apache 2.0; runs in browsers and on laptops. capabilities: [open_weights] placement_confidence: high notable: > OpenAI's first true open-weights model since GPT-2 — signals that narrow security primitives can ship openly while frontier models stay closed. source_refs: [src-openai-privacy-filter] # ---- Poolside Laguna (vertically integrated coding-agent stack) ---- - id: laguna-m-1 name: Laguna M.1 vendor: Poolside release_date: "2026-04-28" year_bucket: 2026 branch: mixture_of_experts secondary_tags: [closed, mixture_of_experts, agentic, code_specialist, us] parents: [] influences: [] openness: closed parameters: "225B total / 23B active" context: "long" modalities: [text] architecture_notes: > Sigmoid MoE router with auxiliary-loss-free load balancing; per-layer head counts allowing decoder layers to vary query-head counts while sharing KV cache shape. Trained from scratch on 30T tokens across 6,144 H200 GPUs in Poolside's in-house Model Factory. capabilities: [tools, agentic] placement_confidence: medium notable: > 46.9 SWE-Bench Pro / 40.7 Terminal-Bench 2.0; shipped alongside the 'pool' terminal agent and 'Shimmer' cloud IDE — the unit of competition is now agent runtime, not weights alone. source_refs: [src-poolside-laguna] - id: laguna-xs-2 name: Laguna XS.2 vendor: Poolside release_date: "2026-04-28" year_bucket: 2026 branch: mixture_of_experts secondary_tags: [open_weights, mixture_of_experts, agentic, code_specialist, edge, small, us] parents: [laguna-m-1] influences: [] openness: open_weights parameters: "33B total / 3B active" context: "long" modalities: [text] architecture_notes: > Same sigmoid-MoE / per-layer-heads architecture as Laguna M.1; runs on a single GPU. Apache 2.0 weights on Hugging Face with full Transformers v5.7.0 support. capabilities: [tools, agentic, open_weights] placement_confidence: high notable: > 44.5 SWE-Bench Pro at one-eighth the active params of M.1 — narrowest open-weights coding gap to a flagship sibling currently on file. source_refs: [src-poolside-laguna] # ---- xAI Grok 4.3 (closed cost-tier agent runtime) ---- - id: grok-4-3 name: Grok 4.3 vendor: xAI release_date: "2026-04-30" year_bucket: 2026 branch: reasoning secondary_tags: [closed, agentic, frontier, tool_using, ultra_long_context] parents: [grok-4-20] influences: [] openness: closed parameters: "Undisclosed" context: "1M" modalities: [text] architecture_notes: > Closed architecture; xAI emphasized agent-tool integration with explicit per-call tool pricing (Web/X Search, Code Execution, File Attachments, Collections Search) rather than reasoning headline gains. capabilities: [tools, agentic, long_context] placement_confidence: medium notable: > Priced at $1.25/$2.50 per M tokens with 207 tok/s output — cost-efficient agent runtime, not frontier intelligence (AA Index 53 vs 60 for GPT-5.5). source_refs: [src-xai-grok-4-3] # ---- W19 (May 8, 2026) additions ---- # ---- OpenAI GPT-5.5 Instant (ChatGPT default refresh) ---- - id: gpt-5-5-instant name: GPT-5.5 Instant vendor: OpenAI release_date: "2026-05-05" year_bucket: 2026 branch: reasoning secondary_tags: [closed, frontier, multimodal, tool_using, mcp] parents: [gpt-5-5] influences: [] openness: closed parameters: "Undisclosed" context: "Undisclosed" modalities: [text, image] architecture_notes: > Tuned light-reasoning variant that replaces GPT-5.3 Instant as the ChatGPT default. OpenAI did not disclose architecture changes; the headline shift is post-training discipline that cuts hallucinated claims by 52.5% on high-stakes prompts (medicine, law, finance). capabilities: [reasoning, tools, mcp, multimodal] placement_confidence: high notable: > 52.5% fewer hallucinations on high-stakes prompts vs GPT-5.3 Instant; AIME 2025 81.2 (vs 65.4); MMMU-Pro 76.0. The 'cheap fallback' tier just inherited frontier-grade math. source_refs: [src-openai-gpt-5-5-instant] # ---- OpenAI GPT-Realtime trio (voice frontier, May 7) ---- - id: gpt-realtime-2 name: GPT-Realtime-2 vendor: OpenAI release_date: "2026-05-07" year_bucket: 2026 branch: multimodal secondary_tags: [closed, frontier, audio, agentic, tool_using, long_context] parents: [] influences: [gpt-5-5] openness: closed parameters: "Undisclosed" context: "128K" modalities: [text, audio] architecture_notes: > First voice model with GPT-5-class reasoning baked in; exposes the same minimal/low/medium/high/xhigh effort knobs as GPT-5.5. Realtime API GA replaces beta; rebuilt WebRTC stack with stateless relay and transceiver split serving 900M weekly users. capabilities: [multimodal, reasoning, tools, agentic, long_context] placement_confidence: high notable: > Big Bench Audio 96.6% (high effort, +15.2pp vs v1.5); Audio MultiChallenge 48.5% xhigh (+13.8pp). Voice crossed into chain-of-thought territory — 'voice model' is now its own tier with Big Bench Audio / Speech Arena leaderboards. source_refs: [src-openai-realtime-2] - id: gpt-realtime-translate name: GPT-Realtime-Translate vendor: OpenAI release_date: "2026-05-07" year_bucket: 2026 branch: multimodal secondary_tags: [closed, audio, multilingual] parents: [gpt-realtime-2] influences: [] openness: closed parameters: "Undisclosed" context: "Undisclosed" modalities: [text, audio] architecture_notes: > Voice-to-voice translation across 70+ input languages and 13 output languages; priced at $0.034 per minute. Sibling of GPT-Realtime-2 specialized for translation workloads. capabilities: [multimodal] placement_confidence: high notable: > BolnaAI independent eval shows 12.5% lower WER on Hindi / Tamil / Telugu vs prior generation. source_refs: [src-openai-realtime-2] - id: gpt-realtime-whisper name: GPT-Realtime-Whisper vendor: OpenAI release_date: "2026-05-07" year_bucket: 2026 branch: multimodal secondary_tags: [closed, audio, edge] parents: [] influences: [gpt-realtime-2] openness: closed parameters: "Undisclosed" context: "Undisclosed" modalities: [text, audio] architecture_notes: > Streaming speech-to-text successor to the Whisper line; lower latency than Whisper-2, priced at $0.017 per minute. Released simultaneously with GPT-Realtime-2 and -Translate as the OpenAI voice trio. capabilities: [multimodal] placement_confidence: high notable: > Streaming-native STT at half the per-minute price of the Translate sibling — completes the OpenAI voice trio for developers building real-time voice agents. source_refs: [src-openai-realtime-2] # ---- Inworld Realtime TTS-2 (frontier voice, conversational turn-awareness) ---- - id: inworld-realtime-tts-2 name: Inworld Realtime TTS-2 vendor: Inworld release_date: "2026-05-05" year_bucket: 2026 branch: multimodal secondary_tags: [closed, audio, multilingual, agentic] parents: [] influences: [] openness: closed parameters: "Undisclosed" context: "Undisclosed" modalities: [text, audio] architecture_notes: > Closed-weights voice model with conversational turn-awareness, emotional steering via natural-language directives ('[whisper]', '[hold back rage]'), and one voice across 100+ languages. Research preview pricing via API. capabilities: [multimodal] placement_confidence: medium notable: > Inworld TTS line holds #1 on Artificial Analysis Speech Arena (Elo 1209.6, 73.3% win rate). TTS-2 extends into real-time dialog — the second 'frontier voice' release of the same week as the OpenAI Realtime trio. source_refs: [src-inworld-tts-2] # ---- Zyphra ZAYA1-8B (open reasoning at 700M active, AMD-trained end-to-end) ---- - id: zaya1-8b name: ZAYA1-8B vendor: Zyphra release_date: "2026-05-06" year_bucket: 2026 branch: reasoning secondary_tags: [open_weights, mixture_of_experts, frontier, small, edge] parents: [] influences: [deepseek-v4-pro, mimo-v2-5-pro] openness: open_weights parameters: "8B total / 700M active" context: "32K" modalities: [text] architecture_notes: > MoE++ lineage with Compressed Convolutional Group-Query Attention (CCGQA) for 8x KV-cache reduction and ~1.7x prefill latency improvement at 16K context. First large-scale frontier-class MoE trained end-to-end on AMD MI300X + Pensando + ROCm via IBM Cloud. Markovian Recursive Self-Attention (RSA) at test time. capabilities: [reasoning, open_weights] placement_confidence: high notable: > AIME 2025 91.9% / HMMT 2025 89.6% with Markovian RSA test-time compute; matches DeepSeek-R1-0528 single-rollout at <1B active — and proves AMD is now a credible frontier-training substrate. source_refs: [src-zyphra-zaya1] # ---- Allen AI EMO (emergent semantic-expert MoE, research-grade) ---- - id: emo-1b14b name: EMO (Emergent Mixture-of-Experts) vendor: Allen Institute for AI release_date: "2026-05-08" year_bucket: 2026 branch: mixture_of_experts secondary_tags: [open_source, research_only, small] parents: [] influences: [] openness: open_source parameters: "14B total / 1B active (8 of 128 experts)" context: "Undisclosed" modalities: [text] architecture_notes: > Research MoE that constrains same-document tokens to share an expert pool, producing experts that specialize at the semantic level (math, code, biomedical) rather than the syntactic level (punctuation). Full weights, code, and data viz published. capabilities: [open_weights] placement_confidence: medium notable: > Maintains near-full performance with only 12.5% of experts active (3pp drop) — the line of research that will set the next 12 months of MoE serving economics. source_refs: [src-allenai-emo] # ---- W20 (May 17, 2026, first Sunday-cadence publish) additions ---- # ---- OpenBMB MiniCPM-V 4.6 (on-device multimodal SLM) ---- - id: minicpm-v-4-6 name: MiniCPM-V 4.6 1.3B vendor: OpenBMB release_date: "2026-05-11" year_bucket: 2026 branch: multimodal secondary_tags: [open_weights, multimodal, edge, small, china] parents: [] influences: [] openness: open_weights parameters: "1.3B (SigLIP2-400M vision + Qwen3.5-0.8B LLM backbone)" context: "262K" modalities: [text, image, video] architecture_notes: > Built on LLaVA-UHD v4 with mixed 4x / 16x visual token compression, cutting visual encoding FLOPs by 50%+ vs prior MiniCPM-V; ships GGUF / BNB / AWQ / GPTQ quantizations across iOS, Android, and HarmonyOS on day one. capabilities: [multimodal, long_context, open_weights] placement_confidence: high notable: > Hits AA Intelligence Index 13 — beats Qwen3.5-0.8B at 19x lower token cost; ships native deployment for iOS / Android / HarmonyOS, collapsing the friction between open-weights release and phone build. source_refs: [src-openbmb-minicpm-v-4-6] # ---- Perceptron Mk1 (specialist video / embodied, closed, sub-Flash-Lite cost) ---- - id: perceptron-mk1 name: Perceptron Mk1 vendor: Perceptron release_date: "2026-05-12" year_bucket: 2026 branch: multimodal secondary_tags: [closed, multimodal, frontier, video, us] parents: [] influences: [] openness: closed parameters: "Undisclosed" context: "32K (multimodal)" modalities: [text, image, video] architecture_notes: > Hybrid reasoning model with structured spatial primitives (point, box, polygon, track, clip) as first-class outputs alongside text. Dynamic 2 FPS video frame rate with temporal grounding. First closed-source release from Perceptron; supersedes the open Isaac series. capabilities: [multimodal, reasoning, tools, agentic] placement_confidence: high notable: > Matches Gemini Pro on video / embodied benchmarks at $0.15 / $1.50 per Mtok — cheaper than Gemini Flash Lite; 85.1 EmbSpatialBench, 72.4 RefSpatialBench. Pricing collapse for physical-AI perception. source_refs: [src-perceptron-mk1] # ---- NVIDIA SANA-WM (open-weights world model on a single GPU) ---- - id: sana-wm name: SANA-WM vendor: NVIDIA release_date: "2026-05-15" year_bucket: 2026 branch: multimodal secondary_tags: [open_weights, multimodal, video, ssm_hybrid] parents: [] influences: [] openness: open_weights parameters: "2.6B (Hybrid Linear Diffusion Transformer)" context: "60 sec @ 720p, 6-DoF camera control" modalities: [video] architecture_notes: > Hybrid Linear Attention combining frame-wise Gated DeltaNet with softmax attention; dual-branch 6-DoF camera control; two-stage generation pipeline with long-video refiner. Trained 15 days on 64 H100s across ~213K public video clips with metric-scale pose supervision. capabilities: [multimodal, open_weights] placement_confidence: high notable: > Generates 60-second 720p clips in 34 seconds on a single RTX 5090 with NVFP4 — first open world model that fits production budgets per-GPU, 36x throughput over prior open baselines. source_refs: [src-sana-wm] # ---- W21 (May 23, 2026, first fully Saturday-cadence publish) additions ---- # ---- Google DeepMind Gemini 3.5 Flash (speed-intelligence Pareto frontier) ---- - id: gemini-3-5-flash name: Gemini 3.5 Flash vendor: Google + DeepMind release_date: "2026-05-19" year_bucket: 2026 branch: multimodal secondary_tags: [closed, multimodal, frontier, agentic, ultra_long_context] parents: [] influences: [] openness: closed parameters: "Undisclosed" context: "1M" modalities: [text, image, audio, video] architecture_notes: > Frontier-class speed-tier model co-developed with the Antigravity 2.0 agent harness. Outperforms Gemini 3.1 Pro across most agentic benchmarks while running at ~280 output tokens per second — ~4x faster than other frontier models. capabilities: [tools, reasoning, agentic, long_context, multimodal] placement_confidence: high notable: > Speed-intelligence Pareto leader at I/O 2026: Terminal-Bench 2.1 76.2%, MCP Atlas 83.6%, GDPval-AA 1656 Elo, AA Intelligence Index 55.3. Pricing $1.50/$9 per Mtok (3x prior Flash list). source_refs: [src-gemini-3-5-flash] # ---- Google DeepMind Gemini Omni Flash (native multimodal video output) ---- - id: gemini-omni-flash name: Gemini Omni Flash vendor: Google + DeepMind release_date: "2026-05-19" year_bucket: 2026 branch: multimodal secondary_tags: [closed, multimodal, frontier, video] parents: [gemini-3-5-flash] influences: [] openness: closed parameters: "Undisclosed" context: "Undisclosed" modalities: [text, image, audio, video] architecture_notes: > First model in the Gemini Omni family — natively multimodal create-and-edit model that reasons across mixed text / image / audio / video inputs and generates grounded video output (10s clips at launch). Built to extend Nano Banana's image paradigm into video; image and audio outputs slated to follow. capabilities: [multimodal] placement_confidence: high notable: > Avatar insertion, cinematic editing via prompt; available immediately in Gemini app, Flow, YouTube Shorts and Create; API in 'coming weeks'. source_refs: [src-gemini-omni-flash] # ---- Alibaba Qwen3.7-Max (first Chinese model in AA Index top 5) ---- - id: qwen-3-7-max name: Qwen3.7-Max vendor: Alibaba release_date: "2026-05-20" year_bucket: 2026 branch: reasoning secondary_tags: [closed, frontier, agentic, china, ultra_long_context] parents: [qwen-3-6-plus] influences: [] openness: closed parameters: "Undisclosed" context: "1M" modalities: [text] architecture_notes: > Text-only proprietary flagship engineered for long-horizon agentic workloads — sustained multi-step operations with tool use, extended-thinking mode, and 1M context (doubled from Qwen3.6-Max's 256K). Sibling Qwen3.7-Plus-Preview adds multimodal. Runs on Alibaba's homegrown Zhenwu M890 silicon (3x predecessor). capabilities: [tools, reasoning, agentic, long_context] placement_confidence: high notable: > First Chinese model in AA Intelligence Index top 5 (56.6, ahead of Gemini 3.5 Flash at 55.3); demonstrated 35-hour autonomous tool-use run with 1,158 calls. source_refs: [src-qwen-3-7-max] # ---- Cohere Command A+ (first Apache 2.0 frontier-adjacent MoE) ---- - id: cohere-command-a-plus name: Command A+ vendor: Cohere release_date: "2026-05-20" year_bucket: 2026 branch: mixture_of_experts secondary_tags: [open_source, mixture_of_experts, multilingual, enterprise, rag_native] parents: [] influences: [] openness: open_source parameters: "218B total / 25B active (128 experts, 8 active)" context: "128K" modalities: [text, image] architecture_notes: > Sparse MoE decoder. First-ever fully Apache-2.0 model from Cohere. Unifies prior Command A, Command A Reasoning, Command A Vision, and Command A Translate into a single set of weights. W4A4 quantization (NVFP4 on experts only with QAD post-training) runs on 1x B200 or 2x H100s. capabilities: [tools, reasoning, agentic, multimodal, open_weights, long_context] placement_confidence: high notable: > First Apache 2.0 frontier-adjacent MoE from a Western lab. τ²-Bench Telecom jumped 37% to 85%; Terminal-Bench Hard 3% to 25%; AA-Omniscience Non-Hallucination #1 at 86%; 48 languages with citation grounding for sovereign / on-prem RAG. source_refs: [src-cohere-command-a-plus] # ---- Microsoft Research Fara1.5-27B (open-weight browser-agent specialist) ---- - id: fara-1-5-27b name: Fara1.5-27B vendor: Microsoft release_date: "2026-05-22" year_bucket: 2026 branch: multimodal secondary_tags: [open_weights, multimodal, computer_use, us] parents: [qwen-3-5] influences: [] openness: open_weights parameters: "27B (dense, Qwen3.5 fine-tune)" context: "Undisclosed" modalities: [text, image] architecture_notes: > Fine-tuned from Qwen3.5 base for browser computer-use. Released as a family (4B / 9B / 27B) integrated with MagenticLite sandboxed browser. Synthetic training data from FaraGen1.5 pipeline using six FaraEnvs app clones; pauses for user approval on ambiguous or irreversible actions. capabilities: [tools, agentic, multimodal, computer_use, open_weights] placement_confidence: high notable: > Online-Mind2Web 72% (beats OpenAI Operator 58.3%, Gemini 2.5 Computer Use 57.3%, Yutori Navigator n1 64.7%); WebVoyager 88.6%. Collapses browser-agent fleet cost structure. source_refs: [src-microsoft-fara-1-5] # ---- Tencent Hunyuan Hy-MT2 (open-weight translation MoE, on-device) ---- - id: hy-mt2-30b-a3b name: Hunyuan Hy-MT2-30B-A3B vendor: Tencent release_date: "2026-05-21" year_bucket: 2026 branch: mixture_of_experts secondary_tags: [open_source, mixture_of_experts, multilingual, china, small] parents: [] influences: [] openness: open_source parameters: "30B total / 3B active" context: "Undisclosed" modalities: [text] architecture_notes: > Flagship of the Hy-MT2 family (1.8B / 7B / 30B-A3B) — fast-thinking MoE translation models supporting 33 languages plus 5 Chinese dialect / ethnic variants. Sibling 1.8B variant uses AngelSlim 1.25-bit quantization to fit in 440MB for on-device inference. Open-source via GitHub, HuggingFace, ModelScope. capabilities: [multilingual, open_weights] placement_confidence: medium notable: > Outperforms DeepSeek-V4-Pro and Kimi K2.6 in fast-thinking translation mode; FLORES-200 leader among open models; powers new Tencent Hy-Translate mini-program with offline mode. source_refs: [src-tencent-hy-mt2] # ---- W23 efficient/local agent substrate additions ---- - id: mellum2 name: Mellum2 vendor: JetBrains release_date: "2026-06-01" year_bucket: 2026 branch: mixture_of_experts secondary_tags: [open_source, mixture_of_experts, code_specialist, agentic, edge] parents: [] influences: [] openness: open_source parameters: "12B total / 2.5B active" context: "Undisclosed" modalities: [text] architecture_notes: > Sparse MoE trained from scratch on natural language and code. Optimized for low-latency, high-throughput text/code workloads: routing, RAG, summarization, validation, sub-agents, and private deployments. capabilities: [tools, agentic, open_weights] placement_confidence: medium notable: > Released under Apache 2.0; JetBrains claims competitive benchmark performance with more than 2x faster inference than similar-size models. Important as efficient agent substrate rather than frontier replacement. source_refs: [src-jetbrains-mellum2] - id: cosmos-3-nano name: Cosmos 3 Nano vendor: NVIDIA release_date: "2026-06-01" year_bucket: 2026 branch: multimodal secondary_tags: [open_weights, physical_ai, world_model, multimodal, synthetic_data] parents: [sana-wm] influences: [] openness: open_weights parameters: "16B (8B reasoner + 8B generator)" context: "Undisclosed" modalities: [text, image, video] architecture_notes: > Open physical-AI omni-model combining world generation, physical reasoning, and action generation. Nano targets workstation-grade inference; Super scales to Hopper/Blackwell research workloads. capabilities: [multimodal, reasoning, open_weights] placement_confidence: medium notable: > Creates an open branch for physical AI reasoning and synthetic-data generation rather than general chatbot competition. source_refs: [src-nvidia-cosmos-3] - id: holo-3-1 name: Holo3.1 vendor: H Company release_date: "2026-06-02" year_bucket: 2026 branch: multimodal secondary_tags: [open_weights, computer_use, agentic, edge, quantized] parents: [fara-1-5-27b] influences: [] openness: open_weights parameters: "0.8B / 4B / 9B / 35B-A3B" context: "Undisclosed" modalities: [text, image] architecture_notes: > Computer-use agent family optimized for web, desktop, mobile, agent frameworks, and local deployment targets. Ships FP8, Q4 GGUF, and NVFP4 checkpoints, including small local variants. capabilities: [tools, agentic, multimodal, computer_use, open_weights] placement_confidence: medium notable: > Moves computer-use agents toward private/local deployment with multiple model sizes and quantized checkpoints. source_refs: [src-hcompany-holo-3-1] # ===== W24 (June 13, 2026) additions ===== - id: claude-fable-5 name: Claude Fable 5 vendor: Anthropic release_date: "2026-06-09" year_bucket: 2026 branch: reasoning secondary_tags: [frontier, reasoning, agentic, coding, closed] parents: [claude-opus-4-8] influences: [] openness: closed parameters: "Undisclosed (inferred larger than Opus 4.8)" context: "1M tokens" modalities: [text, image] architecture_notes: > First publicly released "Mythos-class" tier above Opus, with adaptive thinking and an Opus 4.8 fallback. Debuted #1 on the Artificial Analysis Intelligence Index (64.9). The Mythos 5 sibling is the same model with cyber safeguards lifted, restricted to a government program. capabilities: [reasoning, tools, agentic, coding, multimodal] placement_confidence: medium notable: > First known frontier model to be force-disabled by a government export-control directive (Jun 12, 2026), three days after GA. Commerce withdrew the order Jun 30, 2026 and Fable 5 was restored globally Jul 1, 2026; Mythos 5 remains limited to approved US organizations. source_refs: [src-anthropic-fable-5, src-anthropic-fable-5-redeploy] - id: diffusiongemma name: DiffusionGemma vendor: Google + DeepMind release_date: "2026-06-10" year_bucket: 2026 branch: mixture_of_experts secondary_tags: [open_weights, diffusion, multimodal, edge] parents: [gemma-4] influences: [] openness: open_weights parameters: "25.2B total / ~3.8B active" context: "256K tokens" modalities: [text, image] architecture_notes: > Experimental open text-diffusion model on a Gemma 4 MoE backbone with a block-diffusion head (multimodal in, text out). Generates 256-token blocks per pass at ~1,000+ tok/s on an H100 and ships native vLLM support — a credible open path to low-latency local inference. capabilities: [multimodal, open_weights, reasoning] placement_confidence: medium notable: > First open diffusion LLM with native vLLM support, opening an autoregressive-vs-diffusion architecture split in open weights. source_refs: [src-google-diffusiongemma] - id: north-mini-code name: North Mini Code vendor: Cohere release_date: "2026-06-09" year_bucket: 2026 branch: mixture_of_experts secondary_tags: [open_weights, coding, edge] parents: [command-a] influences: [] openness: open_weights parameters: "~30B total / ~3B active" context: "Undisclosed" modalities: [text] architecture_notes: > Small open-weight coding MoE (~3B active) scoring 27.6 on the Artificial Analysis Intelligence Index. Positioned as a cheap self-host candidate for routine code/text sub-agent calls rather than frontier coding. capabilities: [coding, open_weights, tools] placement_confidence: medium notable: > Cheap self-host coding option in the efficient open-weight tier; not frontier-competitive but a low-cost routing node. source_refs: [src-cohere-north-mini-code] # ===== W25 (June 20, 2026) additions ===== - id: glm-5-2 name: GLM-5.2 vendor: Z.AI (Zhipu) release_date: "2026-06-16" year_bucket: 2026 branch: mixture_of_experts secondary_tags: [open_weights, reasoning, coding, agentic, frontier] parents: [glm-5-1] influences: [] openness: open_weights parameters: "~744-753B total / ~40B active" context: "1M tokens" modalities: [text] architecture_notes: > MIT-licensed sparse-attention MoE ("IndexShare") with MTP speculative decoding and 1M context. Independent testing (VentureBeat) reports it beats GPT-5.5 on several long-horizon coding benchmarks at ~1/6 the cost; Artificial Analysis cites it as the leading open-weight model. capabilities: [reasoning, coding, agentic, open_weights, tools] placement_confidence: medium notable: > Truly permissive (MIT, no regional limits) frontier-adjacent open model — took the open-weight lead in-window and pressures closed pricing. source_refs: [src-zai-glm-5-2] - id: minimax-m3 name: MiniMax-M3 vendor: MiniMax release_date: "2026-06-12" year_bucket: 2026 branch: mixture_of_experts secondary_tags: [open_weights, reasoning, multimodal, coding] parents: [minimax-m1] influences: [] openness: open_weights parameters: "428B total / ~23B active" context: "1M tokens" modalities: [text, image] architecture_notes: > Native-multimodal MoE using MiniMax Sparse Attention; weights posted mid-June with an arXiv report (arXiv:2606.13392) validating ~9x/15x prefill/decode efficiency claims in-window. Released under the MiniMax Community License (commercial use requires a separate agreement) — not OSI-open despite downloadable weights. capabilities: [reasoning, multimodal, coding, open_weights] placement_confidence: medium notable: > Frontier-adjacent coding plus genuine 1M context and native multimodality in one checkpoint, but gated by a non-OSI license. source_refs: [src-minimax-m3] # ===== W27 (July 4, 2026) additions ===== - id: claude-sonnet-5 name: Claude Sonnet 5 vendor: Anthropic release_date: "2026-06-30" year_bucket: 2026 branch: reasoning secondary_tags: [closed, reasoning, agentic, coding, ultra_long_context] parents: [claude-sonnet-4-6] influences: [claude-opus-4-8] openness: closed parameters: "Undisclosed" context: "1M tokens" modalities: [text] architecture_notes: > 1M-token context with adjustable effort levels (max effort adds ~6 Artificial Analysis Intelligence Index points over Sonnet 4.6) and autonomous tool use including browser and terminal. A multi-agent BrowseComp configuration scores 86.6 vs 84.7 single-agent. capabilities: [reasoning, tools, agentic, long_context] placement_confidence: high notable: > Default model for Claude Free/Pro at launch; 85.2% SWE-bench Verified and #5 on the AA Intelligence Index (53 at max effort). Independent testing shows it beating Opus 4.8 on agentic knowledge work while costing ~15% more per completed task at max effort — sticker price and cost-per-task formally diverged. source_refs: [src-anthropic-claude-sonnet-5] - id: gpt-5-6-sol name: GPT-5.6 Sol vendor: OpenAI release_date: "2026-06-26" year_bucket: 2026 branch: reasoning secondary_tags: [closed, frontier, reasoning, agentic] parents: [gpt-5-5] influences: [] openness: closed parameters: "Undisclosed" context: "Undisclosed" modalities: [text] architecture_notes: > Flagship of a three-tier GPT-5.6 family (Sol flagship / Terra balanced / Luna fast) with Programmatic Tool Calling and an "ultra" setting that runs four agents in parallel (Terminal-Bench 2.1 rises from 88.8% to 91.9%). Reached full GA on 2026-07-09 after a 12-day government-coordinated restricted preview with CAISI pre-deployment evaluations; all three tiers are rated High (below Critical) in bio/chem and cybersecurity under OpenAI's Preparedness Framework. capabilities: [reasoning, agentic, tools] placement_confidence: high notable: > GA 2026-07-09 at Sol $5/$30, Terra $2.50/$15, Luna $1/$6 per M tokens — Terra lands at half GPT-5.5's rate, confirming the closed-lab repricing cycle. Sol leads the AA Coding Agent Index v1.1 at 80 (+2.8 over Claude Fable 5) but trails on SWE-Bench Pro (64.6% vs 80%). METR flagged the highest detected reward-hacking rate it has publicly tested, leaving Sol's autonomy time-horizon estimate spanning 11.3 to 270+ hours. GPT-5.4 retires 2026-07-23. source_refs: [src-openai-gpt-5-6-sol, src-openai-gpt-5-6-ga] - id: longcat-2-0 name: LongCat-2.0 vendor: Meituan release_date: "2026-06-30" year_bucket: 2026 branch: mixture_of_experts secondary_tags: [open_weights, agentic, coding, china, ultra_long_context] parents: [] influences: [] openness: open_weights parameters: "1.6T total / ~33-56B dynamically active" context: "1M tokens" modalities: [text] architecture_notes: > MIT-licensed MoE with zero-computation experts for dynamic activation, LongCat Sparse Attention (LSA) for native 1M context, and MOPD multi-expert fusion. Pretrained on 35T+ tokens, claimed end-to-end on 50K+ domestic Chinese ASICs with no NVIDIA hardware. capabilities: [agentic, tools, open_weights, long_context] placement_confidence: medium notable: > Unmasked as "Owl Alpha," the stealth model that led OpenRouter developer charts for two months. Full weight availability was still settling at release (repo initially said "coming soon"; INT8/FP8 uploads observed within days); the domestic-ASIC training claim is vendor-asserted. source_refs: [src-meituan-longcat-2-0] - id: leanstral-1-5 name: Leanstral 1.5 vendor: Mistral AI release_date: "2026-06-30" year_bucket: 2026 branch: reasoning secondary_tags: [open_source, mixture_of_experts, math_specialist, code_specialist, agentic] parents: [] influences: [] openness: open_source parameters: "119B total / 6.5B active" context: "256K" modalities: [text] architecture_notes: > Apache 2.0 MoE trained with mid-training + SFT + RL (CISPO); agentic Lean 4 proof loop using compiler feedback and context compaction (one proof ran 2.7M tokens across 22 compactions). Strong test-time scaling: PutnamBench solves rise from 44 at a 50K token budget to 587 at 4M. capabilities: [reasoning, agentic, tools, open_weights] placement_confidence: medium notable: > Open state of the art on formal theorem proving — miniF2F saturated (100%) and PutnamBench 587/672 at ~$4/problem vs an estimated $300+ for frontier brute force. Predecessor Leanstral (March 2026) is not in the tree, so parents are left empty. source_refs: [src-mistral-leanstral-1-5] - id: nemotron-labs-twotower name: Nemotron-Labs-TwoTower vendor: NVIDIA release_date: "2026-07-01" year_bucket: 2026 branch: mixture_of_experts secondary_tags: [open_weights, diffusion, ssm_hybrid, edge] parents: [nemotron-3-nano-omni] influences: [diffusiongemma] openness: open_weights parameters: "30B total / ~3B active (Nemotron-3-Nano-30B-A3B backbone)" context: "Undisclosed" modalities: [text] architecture_notes: > Block-wise diffusion language model built on a frozen autoregressive backbone: a frozen AR context tower (causal, clean tokens) plus a trainable diffusion denoiser tower (bidirectional block attention, cross-attends to context), trained with ~2.1T tokens vs 25T for the backbone. Backbone is a hybrid Mamba-2 / attention / MoE design; 2.42x wall-clock throughput at 98.7% of AR benchmark quality. capabilities: [open_weights] placement_confidence: medium notable: > Establishes the "diffusion head on a frozen AR backbone" retrofit pattern alongside DiffusionGemma — a cheap post-hoc throughput upgrade rather than a rival pretraining paradigm. One checkpoint supports diffusion, mock-AR, and standard AR decoding. source_refs: [src-nvidia-nemotron-twotower] # ===== W28 (July 11, 2026) additions ===== - id: grok-4-5 name: Grok 4.5 vendor: xAI release_date: "2026-07-08" year_bucket: 2026 branch: reasoning secondary_tags: [closed, agentic, coding, tool_using] parents: [grok-4-3] influences: [] openness: closed parameters: "Undisclosed" context: "1M" modalities: [text] architecture_notes: > First xAI model since the SpaceX combination and the Cursor acquisition; co-trained with Cursor and positioned on cost-per-task rather than benchmark supremacy. xAI claims ~4.2x fewer output tokens than Opus 4.8 to resolve SWE-Bench Pro tasks. capabilities: [reasoning, agentic, tools, coding] placement_confidence: medium notable: > An "Opus-class" workhorse at $2/$6 per M tokens — 4th on the AA Intelligence Index (+16 points over Grok 4.3), 83.3% Terminal-Bench 2.1, but 64.7% SWE-Bench Pro vs Fable 5's 80.4%. Not available in the EU at launch (mid-July target). source_refs: [src-xai-grok-4-5] - id: hy3 name: Hy3 vendor: Tencent release_date: "2026-07-06" year_bucket: 2026 branch: mixture_of_experts secondary_tags: [open_weights, agentic, china, long_context] parents: [hunyuan-2] influences: [glm-5-2] openness: open_weights parameters: "295B total / 21B active (+3.8B MTP layer)" context: "256K (1M via YaRN in community quants)" modalities: [text] architecture_notes: > Apache 2.0 MoE with a 3.8B multi-token-prediction layer usable as a speculative-decoding draft — a llama.cpp PR measured +40% local throughput (24.3 vs 17.4 tok/s) at 91% draft acceptance. Reverses the April preview's regional license exclusions. capabilities: [agentic, tools, open_weights, long_context] placement_confidence: medium notable: > Leads open models on agentic search (84.2 BrowseComp, 91.0 DeepSearchQA, Tencent-reported) at ~$0.20/$0.80 per M tokens, and set the fastest release-to-local pipeline on record for a 295B-class model: community GGUF quants with 1M context landed ~30 hours after release. source_refs: [src-tencent-hy3] - id: muse-spark-1-1 name: Muse Spark 1.1 vendor: Meta release_date: "2026-07-09" year_bucket: 2026 branch: reasoning secondary_tags: [closed, multimodal, agentic] parents: [] influences: [] openness: closed parameters: "Undisclosed" context: "Undisclosed" modalities: [text, image] architecture_notes: > Multimodal reasoning model from Meta Superintelligence Labs with claimed gains in tool use, computer use, coding, and multimodal understanding; runs in "Thinking" mode in the Meta AI app. Muse lineage predecessors are not in the tree, so parents are left empty. capabilities: [reasoning, multimodal, agentic, tools] placement_confidence: low notable: > The structural shift is distribution, not benchmarks: the first frontier Meta model offered through a first-party API (Meta Model API, public preview) rather than open weights or Meta's own apps. source_refs: [src-meta-muse-spark-1-1] # ===== W29 (July 18, 2026) additions ===== - id: kimi-k3 name: Kimi K3 vendor: Moonshot AI release_date: "2026-07-16" year_bucket: 2026 branch: mixture_of_experts secondary_tags: [open_weights, frontier, china, multimodal, agentic, ultra_long_context] parents: [kimi-k2-6] influences: [] openness: open_weights parameters: "2.8T total / 104B active (16 of 896 experts)" context: "1M" modalities: [text, image, video] architecture_notes: > Hybrid 69 Kimi Delta Attention plus 24 Gated MLA layers with Attention Residuals and Stable LatentMoE, MXFP4 quantization-aware training from the SFT stage, and a MoonViT-V2 vision encoder. Weights shipped 2026-07-27 under the Kimi K3 License: MIT-like for internal and most commercial use, but MaaS operators above $20M trailing-twelve-month affiliate revenue need a separate agreement, and products above 100M MAU or $20M monthly revenue owe "Kimi K3" attribution. capabilities: [agentic, tools, multimodal, long_context] placement_confidence: high notable: > Reclassified from closed to open weights on 2026-07-27 when the artifacts and license text landed, and the active-parameter count was corrected from a community estimate of ~50B to a confirmed 104B. The first open-weight-lab model to top an LMArena flagship board (#1 Frontend Code Arena at 1,679 prelim), sitting at 57 on the AA Intelligence Index against roughly 60 for the closed leaders. The license, not the capability, is now the enterprise gate. source_refs: [src-moonshot-kimi-k3, src-moonshot-kimi-k3-weights] - id: inkling name: Inkling vendor: Thinking Machines Lab release_date: "2026-07-15" year_bucket: 2026 branch: mixture_of_experts secondary_tags: [open_weights, frontier, us, multimodal, ultra_long_context] parents: [] influences: [] openness: open_weights parameters: "975B total / 41B active" context: "1M" modalities: [text, image, audio] architecture_notes: > Apache 2.0 MoE pretrained on 45T tokens with trimodal input (text/image/audio), speculative MTP layers, and both BF16 (~2TB) and calibrated NVFP4 (~600GB) checkpoints released day-0 with transformers/vLLM/SGLang/llama.cpp support. NVIDIA confirmed training on GB300 NVL72 systems. capabilities: [open_weights, multimodal, long_context] placement_confidence: high notable: > The largest US-origin open-weights model to date, positioned as a fine-tuning and domain-adaptation base (managed Tinker platform for post-training); an Inkling-Small (276B/12B active) was previewed with weights promised after testing. source_refs: [src-thinkingmachines-inkling] # ===== W30 (July 25, 2026) additions ===== - id: claude-opus-5 name: Claude Opus 5 vendor: Anthropic release_date: "2026-07-24" year_bucket: 2026 branch: reasoning secondary_tags: [closed, frontier, adaptive_thinking, agentic, coding, tool_using] parents: [claude-opus-4-8] influences: [claude-fable-5] openness: closed parameters: "Undisclosed" context: "Undisclosed" modalities: [text, image] architecture_notes: > Adaptive thinking with beta mid-conversation tool-definition changes that preserve prompt-cache continuity and beta automatic fallbacks when a safety classifier blocks the preferred model. Fast mode offers roughly 2.5x response speed at twice the standard token price. capabilities: [reasoning, tools, agentic, multimodal] placement_confidence: high notable: > Near-Fable intelligence at $5/$25 per M tokens — unchanged from Opus 4.8 and roughly half Fable 5 pricing. Anthropic reports state-of-the-art Frontier-Bench and GDPval-AA performance for coding and knowledge work, while Opus 5 remains behind Mythos 5 on cyber capability. source_refs: [src-anthropic-claude-opus-5] - id: gemini-3-6-flash name: Gemini 3.6 Flash vendor: Google + DeepMind release_date: "2026-07-21" year_bucket: 2026 branch: multimodal secondary_tags: [closed, multimodal, frontier, agentic, efficient, ultra_long_context] parents: [gemini-3-5-flash] influences: [] openness: closed parameters: "Undisclosed" context: "1M" modalities: [text, image, audio, video] architecture_notes: > Speed-tier multimodal model optimized for agent fleets. Google reports roughly 17% fewer output tokens than Gemini 3.5 Flash on comparable work while preserving the $1.50 input rate and reducing output price. capabilities: [tools, reasoning, agentic, long_context, multimodal] placement_confidence: high notable: > Launched at $1.50/$7.50 per M tokens, versus 3.5 Flash at $1.50/$9. The economic move is lower completed-task token consumption as well as lower output price; independent fleet-level validation remains pending. source_refs: [src-google-gemini-3-6-flash] # ===== W31 (August 1, 2026) additions ===== - id: inkling-small name: Inkling-Small vendor: Thinking Machines Lab release_date: "2026-07-30" year_bucket: 2026 branch: mixture_of_experts secondary_tags: [open_weights, us, multimodal, efficient, agentic, coding] parents: [inkling] influences: [] openness: open_weights parameters: "276B total / 12B active (6 of 256 experts plus 2 shared)" context: "1M vendor-stated (Artificial Analysis lists 256K)" modalities: [text, image, audio] architecture_notes: > Apache 2.0 MoE with hybrid local/global attention and native text/image/audio input, trained partly by on-policy distillation from Inkling followed by coding RL. BF16 checkpoints need roughly 600GB of aggregate VRAM; calibrated NVFP4 drops that to roughly 180GB. vLLM v0.26.0 shipped an Inkling serving stack the same week. capabilities: [open_weights, multimodal, long_context, tools, agentic] placement_confidence: high notable: > Fulfills the Inkling-Small weights promised at Inkling's July launch, and lands at 40 on the Artificial Analysis Intelligence Index against Inkling's 41 with under a third of the parameters — the strongest efficiency result among US open-weights models to date. Vendor-claimed SWE-bench Verified 80.2% and GPQA Diamond 89.5%. Context window is recorded with a conflict: the vendor model card says up to 1M while AA's write-up lists 256K, unresolved without AA methodology detail. source_refs: [src-thinkingmachines-inkling-small] # ===== W32 (August 8, 2026) additions ===== - id: muse-spark-1-2 name: Muse Spark 1.2 vendor: Meta release_date: "2026-08-05" year_bucket: 2026 branch: reasoning secondary_tags: [closed, multimodal, agentic, us] parents: [muse-spark-1-1] influences: [] openness: closed parameters: "Undisclosed" context: "1M" modalities: [text, image] architecture_notes: > Closed multimodal reasoning model from Meta Superintelligence Labs, served through the Meta Model API at $1.25 input / $4.25 output per 1M with $0.15 cached input. A separate contributor tier prices the same model at $0.10 / $0.20 in exchange for permission to train on customer prompts and completions — roughly 92% off input and 95% off output for data rights. Web search grounding is billed at $2.50 per 1,000 queries and there is no long-context premium. capabilities: [reasoning, multimodal, agentic, tools, long_context] placement_confidence: medium notable: > The first frontier release this publication has recorded where measured capability and per-task cost rose together at unchanged list pricing. Artificial Analysis, given pre-release access, measured Intelligence Index 54 (xhigh) against 1.1's 51 and a GDPval-AA v2 Elo of 1631, up 260 points, while cost per Index task rose from $0.29 to $0.40 because input tokens rose ~53% and output ~36%. Index v4.1.1 then moved it a further +2.7 points on a grader change alone — the largest single move of any model from that patch, and nearly as much as its own release delivered. Its AA-Omniscience gain (18 to 22) is driven by abstention rather than knowledge: hallucination fell 38% to 28% while the attempt rate fell 82% to 67% and accuracy fell 41% to 38%. source_refs: [src-aa-muse-spark-1-2] - id: qwen3-8-max name: Qwen3.8-Max vendor: Alibaba release_date: "2026-08-03" year_bucket: 2026 branch: mixture_of_experts secondary_tags: [closed, frontier, china, multimodal, agentic, ultra_long_context] parents: [qwen-3-7-max] influences: [] openness: closed parameters: "Vendor-stated 2.4T total / 95B active — uncorroborated" context: "1M" modalities: [text, image, video] architecture_notes: > Reasoning model with text, image and video input and text output, GA on the API at $2.00 input / $6.00 output per 1M with cache hits at $0.25 (-88%) and a 7:2:1 blended rate of $1.18. Parameter figures are recorded as vendor-stated and uncorroborated: Alibaba's launch blog describes a 2.4T MoE with 95B active parameters, while Artificial Analysis's model page states Alibaba has not disclosed model size or parameter count. capabilities: [reasoning, multimodal, agentic, tools, long_context] placement_confidence: medium notable: > Independent Intelligence Index of 58 with output speed at 77.6 tok/s, placing it inside the frontier cluster on score. The procurement story is verbosity: Artificial Analysis measured 150M output tokens to run the Index against a class median of 66M, labelled the model "very verbose", and spent $1,741.41 evaluating it. Open weights for the Qwen-Max class and a companion Qwen3.8-27B were announced as coming "next week" but had not shipped as of 2026-08-08, and AA's FAQ confirms this model is not open source. source_refs: [src-aa-qwen3-8-max] - id: lfm2-5-2-6b name: LFM2.5-2.6B vendor: Liquid AI release_date: "2026-08-04" year_bucket: 2026 branch: decoder_only secondary_tags: [open_weights, us, efficient, agentic, edge, commercial_use_restricted] parents: [] influences: [] openness: open_weights parameters: "2.6B" context: "128K" modalities: [text] architecture_notes: > Pre-trained on roughly 34T tokens with the vocabulary doubled to 128K by extending the tokenizer in place, plus a dedicated 128K context-extension mid-training phase. Base and post-trained checkpoints both released. Four-stage post-training: SFT, teacher specialization via RLVR per domain, Multi-Domain On-Policy Distillation, then agentic RL run inside real harnesses with GRPO and an outcome-based reward. Runs at 220 tok/s decode on an M5 Max, 113 tok/s on a Ryzen AI Max+ 395, and roughly 30 tok/s on a phone in under 2.5 GB. capabilities: [agentic, tools, efficient] placement_confidence: high notable: > Recorded with a license contradiction rather than a clean open-weight classification. The release page's "Get Started" section states the model is "Open-weight — Download, fine-tune, and deploy without restrictions", while the LFM Open License v1.0 shipped in the same repository states in Section 5(b) that commercial use by a Legal Entity exceeding the $10M annual revenue Threshold is "not licensed under this Agreement". This is stricter than the Kimi K3 terms recorded in W31, which gated commercial hosting above $20M trailing revenue and pointed to a separate commercial agreement; the LFM text names no such path, and Section 11 terminates the license automatically on any non-compliance. source_refs: [src-liquid-lfm2-5-2-6b, src-liquid-lfm-open-license-v1] - id: muse-glimmer-30b name: Muse Glimmer vendor: Meta release_date: "2026-08-10" year_bucket: 2026 branch: multimodal secondary_tags: [open_weights, dense, us, edge, agentic, tool_using, multimodal, long_context] parents: [muse-spark-1-2] influences: [] openness: open_weights parameters: "30B dense (29.6B including vision encoder)" context: "128K" modalities: [text, image] architecture_notes: > Dense multimodal model distilled from Muse Spark, with a perception encoder for image input and text-only output. All parameters activate on every forward pass rather than routing through an expert subset, which is the unusual choice at this size in 2026. Meta shipped BF16 weights, two 4-bit quantizations, a speculative-decoding drafter and the perception encoder together; the 4-bit deployment envelope is roughly 20 GB, targeting a single consumer GPU on a 24-32 GB system. Training data and training code were not published. capabilities: [agentic, tools, multimodal, open_weights, long_context] placement_confidence: high notable: > The license is the release. Every prior Meta open model shipped under a Llama License carrying a 700-million-user gate, a naming rule and an acceptable use policy; Muse Glimmer carries none of them, and Apache 2.0 leaves Meta no lever to withdraw the grant. Artificial Analysis scores it 35 on the Intelligence Index and 44 on the Openness Index, level with DeepSeek V4 Flash (0731) and GLM-5.2. Meta is not serving it on its own API, so every price and latency number a buyer sees comes from a third party rather than from Meta. source_refs: [src-meta-muse-glimmer, src-aa-muse-glimmer] - id: gpt-5-6-cyber name: GPT-5.6-Cyber vendor: OpenAI release_date: "2026-08-10" year_bucket: 2026 branch: reasoning secondary_tags: [closed, gated, us, frontier, agentic, tool_using] parents: [gpt-5-6-sol] influences: [] openness: gated parameters: "Undisclosed" context: "Undisclosed" modalities: [text] architecture_notes: > Purpose-trained on top of GPT-5.6 Sol for vulnerability discovery and exploit-chain construction. Distributed only through the Daybreak Red tier of a restructured access program: Daybreak Blue hands approved defenders general-purpose frontier models with cyber guardrails removed, Red gates the purpose-trained models behind tighter vetting. Hardware security keys are required for login from 2026-09-01. capabilities: [reasoning, agentic, tools] placement_confidence: medium notable: > The first frontier model whose distribution mechanism is the product decision. Reported to answer 95.0% of advanced cyber requests against 1.5% for GPT-5.6 Sol under normal safeguards, and rated High rather than Critical for cyber capability under OpenAI's Preparedness Framework. Listed at $12.50 input / $75 output per 1M tokens, by a wide margin the most expensive member of the GPT-5.6 family against Sol at $5 / $30. Placement confidence is medium and openness is recorded as gated because no first-party OpenAI source was located during research and no architecture detail has been disclosed. source_refs: [src-openai-gpt-5-6-cyber] - id: grok-4-6 name: Grok 4.6 vendor: SpaceXAI release_date: "2026-08-12" year_bucket: 2026 branch: reasoning secondary_tags: [closed, us, frontier, agentic, coding, tool_using, long_horizon] parents: [grok-4-5] influences: [] openness: closed parameters: "Undisclosed" context: "Undisclosed" modalities: [text] architecture_notes: > Explicitly a post-training release rather than a larger base model: the same foundation as Grok 4.5 with a longer supplemental training run, regenerated supervised fine-tuning trajectories, and reinforcement learning inside agentic environments. Shipped five weeks after Grok 4.5. capabilities: [reasoning, agentic, tools, coding] placement_confidence: medium notable: > Scores 61 on the Artificial Analysis Intelligence Index, tying GPT-5.6 Sol Max and trailing Claude Opus 5 (63) and Fable 5 (62), at $2 input / $6 output against Sol's $5 / $30. The interesting number is not the tie but the turn count: Artificial Analysis measured roughly 53 turns and ~0.5B input tokens to resolve long-horizon agentic tasks against roughly 103 turns and ~2.0B tokens for Claude Opus 5 at max settings. Leads GDPval-AA v2 at 1,753 Elo. Terminal-Bench v3.0 at 26% against 34.6% for GPT-5.6 Sol Max is the gap the release did not close. source_refs: [src-xai-grok-4-6] - id: gemini-3-7-flash name: Gemini 3.7 Flash vendor: Google release_date: "2026-08-13" year_bucket: 2026 branch: multimodal secondary_tags: [closed, us, agentic, coding, tool_using, multimodal, ultra_long_context] parents: [gemini-3-6-flash] influences: [] openness: closed parameters: "Undisclosed" context: "1M" modalities: [text, image, audio, video] architecture_notes: > Workhorse tier of the Gemini 3 family, positioned between the deep-reasoning Pro models and the high-throughput Flash-Lite tier. 1,048,576-token context with 65,536 maximum output tokens; text, image, audio and video input with text output. Released three weeks after Gemini 3.6 Flash and immediately made the model behind Gemini Spark. capabilities: [reasoning, agentic, tools, multimodal, long_context] placement_confidence: medium notable: > Priced at $0.75 input / $3.75 output per 1M tokens, half of 3.6 Flash — but Google's own pricing page states this is introductory through 2026-12-31 and that $1.50 / $7.50 applies from 2027-01-01. The doubling is disclosed at launch rather than discovered later, which makes it the cleanest public example of a lab pre-announcing that today's agent unit economics are a promotional rate. Google gave no timeline for the next Pro model, and Pichai declined Pro-cadence questions on the most recent earnings call. source_refs: [src-google-gemini-3-7-flash] - id: glm-5-3 name: GLM-5.3 vendor: Z.AI (Zhipu) release_date: "2026-08-14" year_bucket: 2026 branch: mixture_of_experts secondary_tags: [china, frontier, agentic, coding, tool_using, long_horizon, ultra_long_context, open_weights] parents: [glm-5-2] influences: [] openness: open_weights parameters: "753B total (architecture identical to GLM-5.2)" context: "1M" modalities: [text] architecture_notes: > Architecturally identical to GLM-5.2 — same 753B mixture-of-experts design, same 1M context. The entire delta is post-training: Z.ai ran the model inside sandboxes built to mimic developer workstations and set it coding tasks, some of which took days to complete, explicitly to train long-horizon behavior. capabilities: [reasoning, agentic, tools, coding, long_context] placement_confidence: medium notable: > GLM-5.3-Flash MIT weights shipped 2026-08-26 confirming Ox Alpha stealth listing, and the full 753B GLM-5.3 checkpoint followed on Hugging Face at zai-org/GLM-5.3 on 2026-08-27/28, closing the two-week weights promise. The gate is now the license, not availability: GLM-5.3 ships under a bespoke GLM-5.3 license rather than MIT, requiring Z.AI security review before commercial use by any Model-as-a-Service operator with group revenue above $10B over any 12 months. Recorded open_weights because the weights are downloadable; the license restriction is noted here rather than by tagging the row gated. source_refs: [src-zai-glm-5-3, src-zai-glm-5-3-flash, src-zai-glm-5-3-hf-weights] - id: ornith-1-5-397b name: Ornith-1.5 vendor: Ornith AI release_date: "2026-08-19" year_bucket: 2026 branch: mixture_of_experts secondary_tags: [open_weights, frontier, agentic, coding, tool_using, long_context] parents: [] influences: [] openness: open_weights parameters: "397B total (35B active); companion 9B distilled variant" context: "256K" modalities: [text] architecture_notes: > Mixture-of-experts flagship released alongside a 35B dense variant and a 9B distilled sibling, licensed MIT. Vendor characterizes the training objective as long-horizon coding and terminal-shell trajectories, with the frontier lineup positioned against DeepSeek V4 Pro and GLM-5.3. No independent tracker (Artificial Analysis, LMSYS) had ranked the model inside the W34 in-window window; all published scores are vendor-run. capabilities: [reasoning, agentic, tools, coding, long_context, open_weights] placement_confidence: medium notable: > Recorded at medium placement confidence because every headline score is vendor-reported and no independent reproduction was available at publish. Ornith AI publishes the Terminal-Bench 2.1 result of 89.7% on its own harness; that harness has not been audited by a third party and the tree does not treat the number as comparable to Artificial Analysis or Vals results until it is. New lab under the DeepReinforce umbrella; no prior model on the tree. Reclassify to placement_confidence: high when an independent benchmark result on a public leaderboard exists. source_refs: [src-ornith-1-5] - id: glm-5-3-flash name: GLM-5.3-Flash vendor: Z.AI (Zhipu) release_date: "2026-08-26" year_bucket: 2026 branch: multimodal secondary_tags: [open_weights, china, frontier, agentic, coding, tool_using, ultra_long_context, mixture_of_experts] parents: [glm-5-3] influences: [] openness: open_weights parameters: "320B total / 18B active per token" context: "1M" modalities: [text, image] architecture_notes: > Natively multimodal MoE released MIT-licensed on Hugging Face after a six-day anonymous Ox Alpha stealth run on OpenRouter. Z.ai reports Artificial Analysis Intelligence Index v4.1.1 score of 57 at $0.045 per task, API list pricing of $0.15/$0.50 per million input/output tokens, and vendor-run coding benchmarks including 84.3% Terminal-Bench 2.1. capabilities: [reasoning, agentic, tools, coding, multimodal, long_context, open_weights] placement_confidence: high notable: > Confirms the Ox Alpha stealth listing (Aug 20–26) as the same checkpoint. Steady-state economics remain unproven: list API carries a 50% launch promo through Sep 9, 2026, and Artificial Analysis Intelligence Index 57 sits below GLM-5.3 (60) and closed frontier tiers. Weights are datacenter-scale MoE, not edge-deployable; no graded source documents the serving hardware behind the stealth run. source_refs: [src-zai-glm-5-3-flash] - id: qwen3-8-flash-next name: Qwen3.8-Flash-Next vendor: Alibaba release_date: "2026-08-26" year_bucket: 2026 branch: mixture_of_experts secondary_tags: [open_weights, china, frontier, agentic, multimodal, ultra_long_context, gated] parents: [qwen3-8-max] influences: [] openness: open_weights parameters: "125B main / 6B active; 51B n-gram embedding params" context: "262K native, 1M via YaRN" modalities: [text, image] architecture_notes: > First open-weight preview of the upcoming Qwen4 hybrid Gated DeltaNet plus Gated Attention architecture. Multimodal MoE with 512 experts and 10+1 active routing; roughly 360GB full-precision checkpoint with an official FP8 variant. Production API SKU Qwen3.8-Flash (separate hosted endpoint at $0.16/$0.47 per million tokens with 1M default context) ships different tooling than the self-hosted artifact. capabilities: [agentic, tools, multimodal, long_context, open_weights] placement_confidence: medium notable: > Architecture preview under qwen-community-1.0 license rather than Apache 2.0. Procurement must treat open checkpoint and production API as different SKUs — the hosted API ships 1M context and official tool calling; Alibaba's release post does not publish a cached-input rate. source_refs: [src-qwen3-8-flash-next] - id: granite-4-2-30b name: IBM Granite 4.2 vendor: IBM release_date: "2026-08-25" year_bucket: 2026 branch: reasoning secondary_tags: [open_weights, enterprise, agentic, coding, tool_using, long_context] parents: [granite-4-0] influences: [] openness: open_weights parameters: "3B, 8B, and 30B dense sizes" context: "512K extended" modalities: [text] architecture_notes: > Apache 2.0 dense family with switchable chain-of-thought mode and native tool calling. 8B and 30B checkpoints add sandbox agentic RL over software engineering, terminal, and web-search environments after roughly 15T token pre-training. IBM reports 57.00% SWE-Bench Verified and 29.24 on Terminal-Bench 2.1 for the 30B variant. capabilities: [reasoning, agentic, tools, coding, long_context, open_weights] placement_confidence: high notable: > Week's clearest enterprise on-prem agent stack without hyperscaler capex exposure — drowned out by Hot Chips but directly substitutable for hosted coding agents when paired with Granite Speech 5.0 Turbo ASR on the same release day. source_refs: [src-granite-4-2] - id: thomson-1-0-small name: Thomson-1.0-Small vendor: Thomson Reuters release_date: "2026-08-24" year_bucket: 2026 branch: mixture_of_experts secondary_tags: [gated, enterprise, legal, long_context] parents: [qwen-3-6-35b-a3b] influences: [] openness: gated parameters: "35B total / 3B active (Qwen3.6-35B-A3B derivative)" context: "262K native" modalities: [text] architecture_notes: > MoE derivative built atop Qwen3.6-35B-A3B, released on Hugging Face under a non-commercial academic license alongside the closed Thomson LLM for CoCounsel tabular legal analysis. Thomson Reuters cites roughly $40M total training investment with a final training run near $450,000 and 35,207 B200 GPU-hours. capabilities: [tools, long_context] placement_confidence: medium notable: > Vertical moat is continual learning on Westlaw and Practical Law corpora, not weight sovereignty — CoCounsel retains multi-model agentic workflows via Anthropic's Claude Agent SDK. Recorded gated because the Hugging Face license is non-commercial academic, not Apache/MIT open weights. source_refs: [src-thomson-1-0-small] - id: hy4-preview name: Tencent Hy4 Preview vendor: Tencent release_date: "2026-08-28" year_bucket: 2026 branch: mixture_of_experts secondary_tags: [open_weights, china, frontier, agentic, coding, ultra_long_context] parents: [] influences: [] openness: open_weights parameters: "770B total / 49B active per token" context: "1M" modalities: [text] architecture_notes: > Apache 2.0 MoE with FP8 variant for 8×GPU serving via official vLLM and SGLang images. Tencent reports a blind internal evaluation of 203 engineering tasks rated by 163 experts scoring 2.99/4.00 versus GLM-5.3 at 2.92 and Kimi K3 at 2.94. OpenRouter lists hosted endpoint at $0.834/$2.501 per million input/output tokens. capabilities: [agentic, tools, coding, long_context, open_weights] placement_confidence: medium notable: > Internal engineering eval is vendor-run; no independent tracker ranking at publish. 31.8% inference throughput gain attributed to self-optimization tooling — treat headline throughput as vendor-stated until reproduced on a neutral harness. source_refs: [src-hy4-preview] - id: gpt-6-astra name: GPT-6 Astra vendor: OpenAI release_date: "2026-09-03" year_bucket: 2026 branch: reasoning secondary_tags: [closed, frontier, multimodal, extended_thinking, agentic, tool_using, computer_use, ultra_long_context] parents: [gpt-5-6-sol] influences: [] openness: closed parameters: "Undisclosed" context: "1M" modalities: [text, image] architecture_notes: > Closed reasoning model released with persistent cross-window memory, asynchronous clarification, computer use, and action monitoring. ARC Prize measured 62.7% on its provider-neutral Standard harness and 99.9% with OpenAI's Provider Adapter, making the released capability explicitly dependent on runtime state preservation and compaction. capabilities: [reasoning, agentic, tools, multimodal, long_context, computer_use] placement_confidence: high notable: > Standard API pricing is $10/$50 per million input/output tokens. The 37.2-point ARC-AGI-3 harness spread is not a model-only comparison; ARC Prize says the closed-ended benchmark is not proof of AGI. source_refs: [src-openai-gpt-6-astra, src-arc-prize-astra] - id: claude-fable-5-1 name: Claude Fable 5.1 vendor: Anthropic release_date: "2026-09-01" year_bucket: 2026 branch: reasoning secondary_tags: [closed, frontier, adaptive_thinking, agentic, code_specialist, tool_using] parents: [claude-fable-5] influences: [] openness: closed parameters: "Undisclosed" context: "Undisclosed" modalities: [text] architecture_notes: > Closed adaptive-reasoning coding model made generally available across GitHub Copilot editor, CLI, coding-agent, web, mobile, and IDE surfaces. Enterprise availability remains administrator-controlled. capabilities: [reasoning, agentic, tools] placement_confidence: medium notable: > GitHub documents 30-day default prompt and output retention for Fable 5 and 5.1 unless an eligible enterprise receives an approved zero-data- retention route; distribution and data policy are part of the SKU. source_refs: [src-github-claude-fable-5-1] - id: gemini-3-8-flash name: Gemini 3.8 Flash vendor: Google release_date: "2026-09-03" year_bucket: 2026 branch: reasoning secondary_tags: [closed, frontier, multimodal, adaptive_thinking, agentic, tool_using] parents: [gemini-3-7-flash] influences: [] openness: closed parameters: "Undisclosed" context: "Undisclosed" modalities: [text, image] architecture_notes: > Closed Flash-tier reasoning model rolled into GitHub Copilot across editor, CLI, cloud-agent, app, and web surfaces. GitHub reports recovery from actionable terminal failures but publishes no reproducible benchmark or exact Copilot multiplier. capabilities: [reasoning, agentic, tools, multimodal] placement_confidence: medium notable: > Introductory provider pricing expires December 31, 2026. Current task economics should not be treated as a defensible 2027 run rate without the post-promotion rate. source_refs: [src-github-gemini-3-8-flash]