Three frontier-scale open checkpoints in 72 hours, and the model card is no longer the number to plan on

Registry movement

Placement notes. Muse Spark Open goes on reasoning rather than mixture_of_experts because Meta positions it as a Spark-family model and the Spark line sits on reasoning; the MoE structure is a secondary tag, as it is for DeepSeek-R1. Beam's five effort levels are the architectural primitive Reflection markets, so it joins reasoning without a parent (a new lab with no prior row). LFM2.5-24B-A2B is an always-think model trained in agent harnesses, which is why it is placed on reasoning rather than beside its dense 2.6B sibling on decoder_only. The vendor on Muse rows is Meta, matching muse-spark-1-2 and muse-glimmer-30b. DeepSeek V4.1 Flash is updated rather than forked because DeepSeek serves 1017 under the same deepseek-flash alias and describes it as a point update; the row now records 763B as the parameter count on the repository's evidence. The medium-confidence flags on muse-spark-open, reflection-beam and muse-wonder record that no harness their vendor did not choose has reproduced the card or that reviewers have withheld a verdict; the flag on lfm2-5-24b-a2b records a licence file not in the packet.

Added

muse-wonder, muse-spark-open, reflection-beam, nemotron-3-5-super-vl, lfm2-5-24b-a2b

Updated

deepseek-v4-1-flash, gemini-4-argon

Google DeepMind · 2026-10-13 · Frontier · Reasoning

Gemini 4 Argon

Generally available to every paid Gemini API developer at an introductory $2 / $10, with a published $4 / $20 reversion and the top LMArena Text score

Teams that held Argon out of their router because nobody outside the Fairwind Program (Google's invitation-only early-access programme) could call it can now put it in, and the decision to make is the date, not the model: the $2 / $10 rate (cached input 95% off, so $0.10 per million cached) runs through November 30 and doubles to $4 / $20, which is Claude Opus 5.5's list price, from December 1, both dates on Google's pricing page. At the introductory rate Artificial Analysis measures $1.99 per index task for a score of 53, level with GPT-6 Astra; at the reverted rate that becomes $3.98, still below Astra on the tracker's own cost-to-run measure. LMArena's October 15 Text board puts Argon first at 1524 ± 9. Google's launch claims (a 2.85x vLLM inference speedup on Gemma 4 E2B, a Chromium refactor across a roughly 6M-line legacy folder, first on the Vals Index) are vendor-reported. The model is not on the free API tier. Budget any Q1 2027 workload at $4 / $20 and treat the next six weeks as the discount.

Model registry ID

gemini-4-argon

Sources Google (The Keyword, developers); Artificial Analysis; LMArena

Meta · 2026-10-12 · Frontier · Reasoning

Muse Wonder

Meta's first frontier-class model prices at Opus 5.5's list and costs 2.8x Spark 1.3 per task to run the independent index

For a buyer the number that matters is not the index score (52 at max effort, 56 for the subscription-only Wonder Full, against 48 for Muse Spark 1.3) but the cost per index task: $4.52 against $1.60 for Spark 1.3, 2.8x, with input tokens per task up 28% (6.03M to 7.69M) and the rest of the increase from the higher rate card; Artificial Analysis calls the cost higher than most similarly intelligent models. A model that reads that much context per task at $4 / $0.20 cached / $20 per million is expensive in exactly the agentic loops it is trained for, and the Active Memory and Active Focus behaviours Meta describes (cross-conversation retrieval; checking in before consequential actions) are model behaviours rather than harness guarantees, which Agent Techniques covers. LMArena's October 15 Text board puts Wonder (Max) second at 1509 behind Argon. Zvi Mowshowitz's 'benchmaxxing' caution and Interconnects' 'vibe checks solid' are the two reviewer positions in circulation, and both are withholding a full verdict; the placement here is medium confidence for that reason. Wonder Full has no per-token API and is deferred on the tree. If you evaluate Wonder, instrument input tokens per task before anything else.

Model registry ID

muse-wonder

Sources Meta AI Research; Artificial Analysis; LMArena

Meta · 2026-10-15 · Open Frontier · Moe

Muse Spark Open

The most capable US open-weight model on the independent index (39), under a licence with an attribution clause and a $20M model-as-a-service gate

Platform teams get an 816B / 42B-active MoE with text, image, audio and video input, a 1,048,576-token context, BF16 and NVFP4 weights, and day-zero llama.cpp, vLLM and SGLang support (SGLang's PR was still open at week's end), with Meta demonstrating Muse Code workflows on two M5 Ultra Mac Studios via MLX. Two things to read before the benchmarks. First, the licence, whose terms the architecture watch carries in full: attribution, an acceptable-use policy that applies to derivatives and outputs (so synthetic-data use inherits it), and a $20M model-as-a-service revenue gate; Interconnects calls it an unnecessary strategic risk, and the tree carries the commercial_use_restricted tag. Second, the independent agentic numbers: 8.1% on Artificial Analysis's Terminal-Bench 4.0 and, per Latent Space's reading of the tables, last among the top nine open models on its four agentic evaluations (outside the graded packet), against Meta-reported Terminal-Bench 2.1 of 76.4 and OSWorld Verified of 80.4 (a benchmark of desktop-computer tasks). The research packet reads the card as sharing Spark 1.3's base; Meta's card does not name the base, and the nine-point index gap to Spark 1.3 (48) is why Interconnects reads it as a safety-tuned Spark 1.2, and why confidence is medium. The Hacker News thread's top-voted comparison was to Qwen3.8 Flash Next and GLM 5.3 Flash, both runnable on a single B300.

Model registry ID

muse-spark-open

Sources Meta AI Research; Hugging Face model card

Reflection AI · 2026-10-16 · Open Frontier · Moe

Reflection Beam 1.0

Beam's Apache 2.0 weights land as promised, and the independent run lands at 30 on the index with 3.0% on Terminal-Bench 4.0

W41 deferred Beam until the repository existed; it now does, with 501B / 23B-active text-only weights in BF16, FP8 and NVFP4, a DFlash2 speculative drafter (a small model that proposes tokens for the large one to verify), day-zero transformers, vLLM and SGLang support, a Rust harness and a Safety Technical Report, so the row is created. The card's final figures rose slightly from the announcement (81.6% SWE-bench Verified, 76.5% SWE-bench Pro v2-Hard, 80.1% Terminal-Bench 2.1, 97.7% AIME 2026, 90.6% GPQA Diamond). Artificial Analysis's run, published the same day (outside the graded packet), scores Beam 30 on the Intelligence Index, 3.0% on Terminal-Bench 4.0 and a 47% hallucination rate on AA-Omniscience (its knowledge-and-abstention test); the 2.1-versus-4.0 difference is unlikely to account for a 77-point gap on its own, but no matched run exists and a broken run (wrong chat template, wrong effort setting) is one of the live explanations, so the card is unreproduced rather than refuted. The FT reported on October 10 that NVIDIA, with $800M invested, is in early talks on a transaction from full acquisition to acqui-hire (outside the graded packet); the licence on the next checkpoint is the thing to watch. For a buyer the Apache 2.0 licence is the cleanest of the week and the independent number is the one to plan on until a matched run exists; Hugging Face is offering $10,000 in Hub credits for community Beam projects, which may produce the third-party runs the card needs.

Model registry ID

reflection-beam

Sources Reflection AI; Hugging Face; Artificial Analysis (outside the graded packet)

DeepSeek AI · 2026-10-17 · Open Frontier · Moe

DeepSeek V4.1-Flash-1017

A point update under the same alias: MIT weights as a 763B checkpoint, Terminal-Bench 4.0 from 31.2 to 42.3 on the card, price unchanged at $0.15 / $0.60

This is the release the week's other open checkpoints are measured against, and it changed nothing about its own price: $0.15 / $0.60 per million off-peak ($0.30 / $1.20 peak), 384K max output, served under the deepseek-flash alias on DeepSeek's API and published on Hugging Face and ModelScope under MIT as a 763B checkpoint in FP8 with FP4 experts (about 510 GB), which settles the 552B-versus-763B discrepancy the W37 row recorded in favour of the repository. The card's Terminal-Bench 4.0 of 42.3 (31.2 for the prior snapshot; DeepSeek's table lists GPT-5.6 Sol at 39.9 and Claude Opus 5 at 51.8) was measured in the Claude Code harness, which is not the harness Artificial Analysis uses, so 'highest published for a downloadable model' compares numbers measured differently; no independent run of 1017 existed at publication and the-decoder notes the same. DeepSWE v1.1 77.0, Automation-Bench 60.5, Agents' Last Exam 35.4 and HLE 42.6 are card figures. Operational notes: the release omits a Jinja chat template (the file that tells a serving stack how to format a conversation) in favour of a reference encoder and a Rust 'deepseek-recipe' toolkit, and third-party NVFP4, MLX and GGUF quantizations (210-432 GB) appeared within hours. The 'DOA' verdict on Muse Spark Open in circulation (@teortaxesTex, via Latent Space) is anchored on this model's price; the independent score is still the 39 of the prior snapshot until the tracker runs 1017.

Model registry ID

deepseek-v4-1-flash

Sources Hugging Face model card; DeepSeek API news

NVIDIA · 2026-10-15 · Open Frontier · Multimodal

Nemotron 3.5 Super VL

A 120B / 12B-active hybrid Mamba-Transformer vision-language MoE under OpenMDW, NVFP4 on day zero, with a 24 GB Bolt derivative

This is the week's clearest example of a model sized to the hardware its publisher sells: NVFP4 checkpoints on day zero, up to 4,856 output tokens per second per GPU on GB200 (vendor figure), a Bolt 74B-A9B derivative compressed to fit 24 GB and measured by NVIDIA at 84 tokens per second on the RTX Spark laptops that went on sale the next day, and day-one support in llama.cpp, Ollama 0.40.5, LM Studio, Unsloth and vLLM. The technical report claims 51.78% on SWE-bench Pro with about 4.6x fewer active parameters than Nemotron 3 Ultra, 71.10% SWE-bench Verified, 66.70% Terminal-Bench 2.1 and 99.62% median NVFP4 accuracy recovery, all vendor-run; Artificial Analysis scores it 26 on the Intelligence Index (outside the graded packet), and the 'up to 5x higher throughput' claim names no comparison model or operating point. OpenMDW-1.1 is a permissive licence with no revenue gate. For an architect deciding what runs on a 128 GB notebook or a single workstation GPU, Bolt is the first in-window artifact that was built for that envelope rather than quantized down to it. The row is placed at high confidence because the architecture, training recipe and licence are fully documented in the technical report, and the confidence flag records how well the tree knows what the artifact is, not how well its card has been reproduced; the scores are labelled vendor-run where they are.

Model registry ID

nemotron-3-5-super-vl

Sources NVIDIA blog; Nemotron 3.5 technical report

Liquid AI · 2026-10-13 · Edge · Moe

LFM2.5-24B-A2B

Liquid's largest LFM2.5: a 24B / 2B-active always-think MoE with 131K context at 184 tokens per second on a laptop, vendor-measured

Edge and on-device teams get the first LFM with a 131K context (from 32K), a 128K vocabulary, a hybrid of gated short convolutions and grouped-query attention under 64 experts with top-4 routing, and base plus post-trained checkpoints with GGUF, MLX, ONNX, vLLM and SGLang support. Liquid's table has it winning five of eight agentic benchmarks against Qwen3.6-35B-A3B (48.09 versus 38.19 on LiveClaw, 27.69 versus 8.72 on AutomationBench) and running at 184 tokens per second on an Apple M5 Max and up to 21K output tokens per second on a single B300; every figure is vendor-run and no independent index placement is in the graded packet. The licence file was not in the packet either, and the 2.6B row's LFM Open License v1.0 carried a $10M revenue gate with automatic termination on non-compliance, so the row is open_weights at medium confidence pending that check; teams above the gate should read the licence before the benchmarks. Liquid's other event this week was d1-ultra taking the Decision Index lead at 65.2 (outside the graded packet), which keeps the decision-model category deferred but moves its leader away from the startup that defined it.

Model registry ID

lfm2-5-24b-a2b

Sources Liquid AI blog; Hugging Face model card

Card versus harness: the open-weight gap moves from downloadability to reproducibility

Last week the open-weight problem was that the label came without an artifact; this week every artifact arrived and the scores came apart instead. The independent Artificial Analysis figures for Beam, Muse Spark Open and Nemotron sit outside the graded packet and are labelled at every use; the vendor figures are from the cards. Four mechanisms could explain the gap and none of them is fraud: Terminal-Bench 2.1 and 4.0 are different task sets; DeepSeek measured inside the Claude Code harness, which supplies tool scaffolding the tracker's harness does not; reasoning-effort settings are rarely matched; and an independent run can itself be broken (wrong chat template, wrong effort level), which for a two-day-old checkpoint is a live possibility. The one same-benchmark reference point, Haiku 5.5's 6.4-point vendor-versus-tracker gap in W41, is a single model and not a distribution, so it bounds nothing; it says only that a harness effect of that size exists. The figures are therefore unreproduced, not refuted. The procurement consequence is the same whatever the mechanism: a model card's agentic row is the vendor's upper bound under the vendor's harness, and the acceptance test for an on-prem deployment is a run in the harness you will actually use. The AI Stack Weekly's pattern watch tracks the label-artifact gap as a seven-week inductive pattern; the confirming condition it names for next week is an independent 1017 run below 35 on Terminal-Bench 4.0 (p138). Until that run exists, the open-frontier scorecard holds the models whose independent scores are known.

Examples

Reflection Beam: card 80.1% Terminal-Bench 2.1 / 81.6% SWE-bench Verified; Artificial Analysis index 30, Terminal-Bench 4.0 3.0%, 47% hallucination (outside the graded packet), Muse Spark Open: card Terminal-Bench 2.1 76.4, OSWorld Verified 80.4; Artificial Analysis Terminal-Bench 4.0 8.1%, last of top nine open models on agentic evals per Latent Space's reading (outside the graded packet), DeepSeek V4.1-Flash-1017: card Terminal-Bench 4.0 42.3 in the Claude Code harness; no independent run yet, Nemotron 3.5 Super VL: card 51.78% SWE-bench Pro; Artificial Analysis index 26 (outside the graded packet), Single reference point: Haiku 5.5 in W41, vendor 39.2% versus independent 32.8% on Terminal-Bench 4.0, a 6.4-point gap on one model

Sources Reflection AI; Meta AI Research; Hugging Face (DeepSeek); Artificial Analysis (outside the graded packet)

Four licences for four open checkpoints, and the restrictive one came from the lab that does not sell silicon

This is the fact home for the week's licence terms; the AI Stack Weekly's synthesis observes that the terms line up with who finances the releaser, and this entry supplies the table it rests on. Two of the three permissive frontier-scale licences this week (Apache 2.0 on Beam, OpenMDW on Nemotron) came from a lab NVIDIA has invested $800M in and from NVIDIA itself; the third (MIT) came from DeepSeek, as every DeepSeek release has. The one restrictive licence came from Meta, whose financier is its own advertising business. For a platform team the operative clauses are concrete: Muse's attribution requirement travels with derivatives, its acceptable-use policy covers outputs (so a post-training set distilled from Muse Spark Open inherits it), and the $20M gate is on the provider's revenue across the preceding 12 months, not revenue from the model; secondary summaries differ on the period (Interconnects writes 'per month', Latent Space 'annual'), and the licence file in the repository governs. Interconnects' observation that Western open models historically out-adopt Chinese models of equal capability is the strongest argument that the licence will be ignored in practice, as Llama's 700-million-user clause largely was; Interconnects' open question is whether the output clause chills synthetic-data use, which is the fastest-growing consumer of open weights. The China draft catalogue is the week's other licence-adjacent fact: it is a supply-side restriction on the open models that currently lead the independent index, it is a draft, and none of this week's releases crosses its thresholds. The tree records the Muse gate with the commercial_use_restricted tag so the filter on /industry/tree shows it. The weekly's open question on NVIDIA and Reflection is the test for the Apache 2.0 leg: if a transaction closes and Beam's next checkpoint is relicensed, the pattern breaks.

Examples

DeepSeek V4.1-Flash-1017: MIT, Reflection Beam 1.0: Apache 2.0 (NVIDIA $800M investor; acquisition talks reported by the FT, outside the graded packet), Nemotron 3.5 Super VL and Bolt: OpenMDW-1.1, Muse Spark Open: Muse Model License ('Built with Muse Spark Open' attribution; acceptable-use policy on derivatives and outputs; separate agreement for model-as-a-service providers with over $20M in revenue across the preceding 12 months, as the Hugging Face card's licence section states it), LFM2.5-24B-A2B: licence file not in the graded packet; the 2.6B sibling's LFM Open License v1.0 gated commercial use above $10M revenue, FLUX 3 Image Klein (excluded from the tree): Kommunity License, commercial output below $5M revenue only, China, October 16 (outside the graded packet): a draft export catalogue proposing licences for exporting model weights, checkpoints and source code above 2T parameters, 10^26 training FLOPs or a global top-10 ranking, with automatic 'report and export' licences for open-sourcing and an exemption for weights already public; DeepSeek's 763B release sits under the thresholds

Sources Hugging Face model cards (Meta, Reflection, DeepSeek); NVIDIA; Liquid AI

NVFP4 as the day-zero checkpoint format, and a serving stack that absorbed five frontier-scale releases in a week

The deployability half of the open-weight story is now measured in hours rather than weeks, and the format the checkpoints arrive in is NVIDIA's. Four of the week's five open releases shipped an NVFP4 checkpoint on day zero, vLLM's 814-commit release added an NVFP4 KV cache, and the NVIDIA-generation numbering of the supported architectures (SM8x, SM90, SM12x) tells an operator which hardware the format was built for; the hyperscaler-silicon and AMD paths run the BF16 or FP8 weights and their own quantizers. The llama.cpp detail is the one for anyone planning a 2-bit deployment: the Muse Spark Open quantizer keeps some layers at 8-bit because the low-bit types caused looping, so the headline Q2_K size understates the memory a working build needs. vLLM's breaking changes (Python 3.10 dropped, default CORS restricted, three GitHub security advisories fixed) are the week's upgrade-gating items for platform teams, and the default-on structured-decision endpoint is the serving layer catching up with the decision-model category the tree still defers. The AI Stack Weekly's hardware lens carries the notebooks and workstation GPUs these artifacts were sized for; the point here is that the serving stack is no longer the bottleneck between a repository and a deployment, which moves the bottleneck to the independent evaluation the first architecture-watch entry describes.

Examples

NVFP4 (NVIDIA's 4-bit floating-point weight format) checkpoints on day zero: Beam, Muse Spark Open, Nemotron 3.5 Super VL and Bolt, FLUX 3 Image Klein; DeepSeek ships FP8 with FP4 experts and third-party NVFP4 within hours, vLLM v0.32.0 (Oct 18, added at the research window's close): 814 commits from 363 contributors, NVFP4 KV cache (the cache of attention state that holds a conversation's context in GPU memory) on SM8x / SM90 / SM12x GPU generations, FP8 KV cache by default for GLM-5.x on SM100, default-on POST /v1/systemone structured-decision endpoint, generic 'hf' tool and reasoning parsers, Python 3.10 dropped, default CORS restricted, llama.cpp b11671: Muse Spark Open text, vision and audio (PR #30616), Q8_0 / Q4_K_S / Q2_K GGUF builds validated against the full-precision model; some layers held at 8-bit because lower-bit versions produced looping output, Ollama 0.40.5 (Nemotron 3.5 Super and Bolt) and 0.40.6 (q2_0 tensors); SGLang PR #44530 (+41,020 lines) still open, Muse Spark Open on two M5 Ultra Mac Studios via MLX; Bolt 74B-A9B at 84 tokens per second on RTX Spark (vendor)

Sources vLLM GitHub releases; llama.cpp PR #30616; Ollama releases; SGLang PR #44530

Introductory rates with published reversions: the $2 / $10 band now has a calendar

This is the fact home for the week's API rate cards (OpenAI's ChatGPT Enterprise per-token card is a seat-plus-usage contract and lives in The Application Layer). Three purchasable models now sit at $2 / $10 (Sonnet 5.5, GPT-6.1 Sol, Argon) and one of them carries a published doubling: Google's pricing page runs the $2 / $10 rate through November 30 and lists $4 / $20 from December 1. That makes Argon's effective price a function of when the workload runs, which is the seventh week in a row that the AI Stack Weekly's pattern watch has recorded a list price moving through a non-list mechanism (dated promotions, multipliers, credit caps). Muse Wonder's $4 / $20 is the first new frontier release in months to launch at Opus 5.5's list with no promotion, and its $4.52 per index task says the list price is not where its cost sits. DeepSeek is the control: a capability update with no price change, at roughly a thirteenth of the band's input price and a sixteenth of its output price. For a router the rule stays what it was in W41, key on prompt-length band and call type, with one addition: carry a reversion date per model, and re-run the routing table on December 1.

Examples

Gemini 4 Argon: $2 / $10 introductory through November 30, cached input 95% off; $4 / $20 from December 1, both dates on Google's pricing page, Muse Wonder: $4 / $0.20 cached / $20, no promotion; Muse Wonder Full subscription-only, DeepSeek V4.1-Flash-1017: $0.15 / $0.60 off-peak, $0.30 / $1.20 peak, unchanged through a capability update, Reference rows from W41: Sonnet 5.5 $2 / $10 with $0.10 cache reads; GPT-6.1 Sol $2 / $10; Mistral Large 4 $0.68 / $2.09 on an undated 50% discount

Sources Gemini API pricing; Meta Developers; DeepSeek API news

The Decision Index flips in 48 hours and the LLMs still win on knowledge

The decision-model category (small models that return a classification or approve/flag decision rather than generated text) got its first independent index this week and the index's own numbers make the architectural argument the tree has been waiting on. A general LLM at its lowest effort setting (Haiku 5.5 at 61) still outscores every decision model at launch on the knowledge-heavy components, at roughly 11x the latency and 4x the cost per million tasks; the decision models win on speed and price, which is a serving property, not an architectural one. That is consistent with the fine-tuning-recipe reading of the 3-of-6 test's architectural-distinction condition, so the category stays deferred. The leadership change is the newer fact: the index debuted with Jev 1.13 at 58 and within 48 hours Liquid's d1-ultra posted 65.2 at the same $0.042 input price (outside the graded packet), with an 85% non-hallucination rate against 35-37% for Jev, while Jev 1.14 scored below 1.13. The graded packet's software-04 still carries the launch-day leader. For an architect routing approve, flag and classify calls, the index gives a first public basis for choosing a decision model on accuracy rather than on price alone, and the first evidence that the cheap tier's ranking is not stable week to week. Artificial Analysis also added FrontierCode 1.1 (leading models fail more than 40% of maintainer-mergeability tasks) and AA-3D the same week; neither moves a scorecard row yet.

Examples

Artificial Analysis Decision Index v1.0 (Oct 13): AA-Omniscience-Decision, HLE-MultiChoice, When2Call, WildGuardTest at 25% each, Launch leader TypeSafe Jev 1.13 at 58; GLiDE and Microsoft-Decision-1 at 54; Perplexity Decider v1.1 27B cheapest at $17 per million tasks, Liquid d1-ultra at 65.2 by the Oct 15-16 snapshot at $0.042 per million input tokens, output free; Jev 1.14 at 57.1, below 1.13 (outside the graded packet), Claude Haiku 5.5 (Low) scores 61 at 1.4 s per task and $184 per million tasks against 0.13 s and $42 for the decision models

Sources Artificial Analysis (Decision Index launch); Artificial Analysis changelog

Artificial Analysis Intelligence Index (independent), closed tier

Two new entries from Meta enter the top five; Argon's score is unchanged by general availability. Setting in parentheses is the reasoning-effort level and scores are not comparable across settings

Claude Opus 5.5

58

Muse Wonder Full (max; subscription-only)

56

Gemini 4 Argon (High)

53

GPT-6 Astra

53

Muse Wonder (max)

52

Sources Artificial Analysis (Muse Wonder and Gemini 4 Argon articles)

Artificial Analysis Intelligence Index (independent), open-weight tier

The best US open score moves from 26 to 39 inside 36 hours and every US entrant finishes behind the Chinese models already on the tree; the Beam, Nemotron and MiMo/GLM rows are Artificial Analysis figures outside the graded packet

MiMo-V2.6-Pro (outside the graded packet)

46

GLM-5.3 (outside the graded packet)

45

Muse Spark Open (Meta-reported; matches Artificial Analysis's listing)

39

Reflection Beam 1.0 (outside the graded packet)

30

Nemotron 3.5 Super VL (outside the graded packet)

26

Sources Meta AI Research; Artificial Analysis (outside the graded packet)

Terminal-Bench, vendor card versus independent harness, with anchors

The card-to-independent gap for the week's open checkpoints is 68-77 points where both exist; the task set (2.1 versus 4.0), the harness (DeepSeek measured in Claude Code) and effort settings differ, nobody has yet shown how much each accounts for, and the figures are unreproduced rather than refuted. The anchor rows show where known models sit on 4.0

Reflection Beam 1.0 (card, Terminal-Bench 2.1 / Artificial Analysis, Terminal-Bench 4.0; the latter outside the graded packet)

80.1% / 3.0%

Muse Spark Open (card, Terminal-Bench 2.1 / Artificial Analysis, Terminal-Bench 4.0; the latter outside the graded packet)

76.4 / 8.1%

DeepSeek V4.1-Flash-1017 (card, Terminal-Bench 4.0, Claude Code harness; prior snapshot 31.2)

42.3

Anchors on Terminal-Bench 4.0, closed models: Muse Wonder Full (Artificial Analysis) / Claude Opus 5 (DeepSeek's card table) / GPT-5.6 Sol (DeepSeek's card table)

57.1 / 51.8 / 39.9

Anchor on Terminal-Bench 4.0, same-benchmark harness gap: Haiku 5.5 vendor / Artificial Analysis (W41; one model, not a distribution)

39.2 / 32.8

Sources Reflection AI; Meta AI Research; Hugging Face (DeepSeek); Artificial Analysis (outside the graded packet)

LMArena Text Arena, October 15 board (independent, preference-based)

Argon takes first on general availability and Muse Wonder debuts second; arena scores measure pairwise human preference, not task completion

Gemini 4 Argon

1524 ± 9

Muse Wonder (Max)

1509

Sources LMArena

Artificial Analysis cost to run the Intelligence Index (independent)

Muse Wonder's per-task cost is higher than most similarly intelligent models (Artificial Analysis's wording) and 2.8x Muse Spark 1.3's $1.60; Argon's cost doubles on reversion; the Muse Spark Open and DeepSeek Flash rows are Artificial Analysis figures outside the graded packet

Muse Wonder (max; 7.69M input tokens per task, up 28% on Spark 1.3)

$4.52 per task

Gemini 4 Argon (High; introductory rate through November 30, then from December 1)

$1.99, rising to $3.98

Muse Spark Open (hosted; outside the graded packet)

$0.84 per task

DeepSeek V4.1 Flash (prior snapshot; API, then OpenRouter; outside the graded packet)

$0.49 / $0.27 per task

Sources Artificial Analysis (Muse Wonder and Gemini 4 Argon articles); open rows outside the graded packet

Artificial Analysis Decision Index v1.0 (independent), launched October 13

The launch leader was displaced within 48 hours; the d1-ultra and Jev 1.14 rows are from the October 15-16 snapshot (outside the graded packet), the rest from the launch article

Liquid d1-ultra (October 15-16 snapshot; outside the graded packet)

65.2

Claude Haiku 5.5 (Low; an LLM, 1.4 s per task, $184 per million tasks)

61

TypeSafe Jev 1.13 (launch leader)

58

TypeSafe Jev 1.14 (October 15-16 snapshot; outside the graded packet)

57.1

GLiDE and Microsoft-Decision-1

54

Sources Artificial Analysis (Decision Index launch article; October 15-16 snapshot outside the graded packet)

2026-10-15 · Meta

Publishes Muse Spark Open under the Muse Model License (terms in the architecture watch) with 'additional safety post-training' on the card; the research packet reads the card as sharing Muse Spark 1.3's base, and Meta's card does not name the base

Meta has reopened a US open-weight line at frontier scale for the first time since Llama 4 and kept its most capable models closed, which Alexandr Wang stated as policy (The Next Web, outside the graded packet); the licence is more restrictive than Muse Glimmer's Apache 2.0 in August. For any provider near the $20M gate or any team building synthetic data from outputs, the licence is the deployment decision, and the unresolved 1.2-versus-1.3 base question is why the tree's row is medium confidence.

Sources Meta AI Research; Hugging Face model card

2026-10-13 · Google DeepMind

Opens Gemini 4 Argon to all paid Gemini API developers and AI Ultra subscribers in AI Studio and Antigravity at $2 / $10 with cached input 95% off through November 30; the pricing page lists $4 / $20 from December 1; not available on the free tier

The thirteen-day Fairwind gate was shorter than the W40 and W41 watchlists expected, and the price is a promotion with a published destination and date rather than a list. Budget holders should model Argon at $4 / $20 for anything running from December 1 and treat the $2 / $10 window as the evaluation period; at the reverted rate Argon is priced level with Opus 5.5 and Muse Wonder.

Sources Google (The Keyword); Gemini API pricing

2026-10-17 · DeepSeek AI

Ships V4.1-Flash-1017 under the existing deepseek-flash alias with MIT weights (763B, about 510 GB), unchanged $0.15 / $0.60 off-peak pricing, a reference encoder in place of a Jinja chat template and a Rust 'deepseek-recipe' toolkit

A capability update with no price change is the behaviour that anchors the 'DOA' verdict on Muse Spark Open in circulation (@teortaxesTex via Latent Space); the alias reuse means API customers were moved to the new snapshot without a model-id change, so platform teams that pin behaviour should re-run their evaluations against the alias and pin to a dated snapshot where the provider offers one. The missing Jinja template is a small operational break for serving stacks that assume one, and the 763B figure resolves the parameter-count discrepancy the tree carried since W37.

Sources DeepSeek API news; Hugging Face model card

2026-10-15 · NVIDIA

Releases Nemotron 3.5 Super VL and Bolt under OpenMDW-1.1 with NVFP4 checkpoints on day zero, one day before RTX Spark laptops (128 GB unified memory, from $2,599) go on sale; the FT reports NVIDIA in early talks to acquire or acqui-hire Reflection AI, whose Apache 2.0 Beam shipped October 16 (FT, outside the graded packet)

NVIDIA is simultaneously the week's second-largest open-weights publisher, the investor behind the third, and the vendor whose number format (NVFP4) four of five open releases shipped in on day zero. For buyers the practical effect is that the open-weight stack is being packaged for NVIDIA hardware first; for the tree the open question is the licence on Beam's next checkpoint if a transaction closes. The 'up to 5x throughput' claim for Super VL names no comparison model or operating point and is carried as vendor-reported.

Sources NVIDIA blog; NVIDIA Newsroom; FT via Yahoo Finance (outside the graded packet)

2026-10-16 · Reflection AI

Delivers Beam 1.0 weights under Apache 2.0 in BF16, FP8 and NVFP4 with a DFlash2 drafter, a Safety Technical Report and a Rust harness, with final card figures slightly above the October 5 announcement; Hugging Face offers $10,000 in Hub credits for community Beam projects

Reflection kept the 'later this month' promise W41 deferred on, which moves it out of the label-without-artifact column; the card figures remain unreproduced, and the independent run that exists is far below them. The Hub-credit programme is the mechanism most likely to produce third-party harness runs, which is what the tree's medium-confidence flag is waiting for. Buyers get the cleanest licence of the week on a model whose independent number is 30.

Sources Reflection AI; Hugging Face blog

2026-10-15 · Mistral AI

Releases Timestral, four open zero-shot time-series forecasters (5M to 209M parameters, 21 quantiles, 32,768-step context, CPU-capable) claiming first place among pretrained entries on GIFT-Eval, TIME and fev-bench from a single checkpoint, with fine-tuning through Mistral Forge; Mistral Large 4's weights remain due at the end of October

Mistral's first non-language family lands as open weights two weeks before its trillion-parameter language model's weights are due, which is the sequence a lab follows when the small release carries no competitive cost and the large one does. Timestral is outside the tree's scope; the Large 4 repository and its licence (the AI Stack Weekly bets neither Apache 2.0 nor MIT by November 15) are the Mistral items that move a row. Leaderboard claims are vendor-reported.

Sources Mistral AI; GitHub mistralai/timestral

Next two weeks

Artificial Analysis publishes Intelligence Index and Terminal-Bench 4.0 runs for DeepSeek V4.1-Flash-1017

The first independent number for the week's most-cited open checkpoint; a Terminal-Bench 4.0 result below 35 (more than seven points under the card's 42.3) is the AI Stack Weekly's p138 and the confirming condition for its label-artifact pattern, and the result decides the sub-dollar and open-frontier rows.

Oct 23

Nemotron 3.5 Bolt E4B 4GB release

The second Bolt derivative, sized for 4 GB; it is reviewed for the edge row and for whether NVIDIA's compression recipe holds at a size where the independent index has no comparable open model.

Oct 27-31

Mistral Large 4 weights and licence on Hugging Face

The tree's closed-at-medium-confidence row flips to open_weights if the files land; the licence resolves the weekly's p132 (neither Apache 2.0 nor MIT by November 15) and adds a fourth data point to the licence-and-financing observation the architecture watch tabulates.

Open; FT reported a decision possible 'within weeks' on October 10

NVIDIA-Reflection transaction terms and the licence on Beam's next checkpoint

A closed transaction with Beam relicensed restrictively would break the week's licence pattern and change the reflection-beam row's openness; a transaction that leaves Apache 2.0 in place confirms it. The talks are single-sourced to the FT and unconfirmed by either company.

Dec 1 (Google's pricing page dates the $2 / $10 rate through November 30)

Gemini 4 Argon introductory pricing reverts to $4 / $20

The first published doubling in the $2 / $10 band; usage share on OpenRouter and the Gemini API before and after is the test of whether a promotional frontier price carries volume through its reversion, and the scorecard's challenger slot is re-examined at the new price.

Open

Independent harness runs for Muse Spark Open and Reflection Beam on any Terminal-Bench version their cards report; Meta's answer on the Spark Open base (1.2 or 1.3)

Both rows are at medium confidence because no harness their vendor did not choose has reproduced the card; a run within five points on the same benchmark version would lift the confidence (and a run that reproduces the gap on a matched version would turn 'unreproduced' into 'refuted'), and Meta's base statement decides the parent on the tree.

Q4 2026

Artificial Analysis Decision Index second snapshot and any OpenAI entry

The index flipped in 48 hours; a stable second snapshot and OpenAI's absence or presence (the Decisions API GA is the weekly's p133) decide whether the decision-model category gets the third technical report the 3-of-6 test needs.

• 2026-10-17: Issue 26 published. Dating convention: the issue is dated Saturday and the research window closes Sunday; vLLM v0.32.0 (October 18) was added at window close and is labelled as such. Five tree rows added (muse-wonder, muse-spark-open, reflection-beam, nemotron-3-5-super-vl, lfm2-5-24b-a2b) and two updated (deepseek-v4-1-flash for the 1017 snapshot and 763B MIT checkpoint; gemini-4-argon from gated to closed and active). Muse Wonder Full, the decision models (Liquid d1-ultra, Jev 1.14), StepFun Step 5 Preview and the Nemotron Bolt derivatives are reviewed and deferred; Timestral, FLUX 3 Image Klein, Muse Code and Muse Spark 1.3, the OpenAI enterprise rate-card SKUs and the serving-stack releases are reviewed and excluded.
• Revision cycle 1 (editorial board): the medium-confidence count corrected from three to four of five additions, with the four named; 'lead changed hands three times' corrected to twice; Muse Wonder's 2.8x restated as cost per task with input tokens per task up 28% (6.03M to 7.69M) and the 'highest per-task cost' superlative replaced with Artificial Analysis's wording; the Spark Open base statement downgraded to the research packet's reading, since Meta's card does not name the base; Google's own dating of Argon's rate (through November 30; $4 / $20 from December 1) replaces every third-party-date qualifier; the Terminal-Bench table gains anchor rows and the card-versus-harness gaps are described as unreproduced rather than refuted, with a broken independent run named as a live explanation; the 'DOA' verdict attributed to @teortaxesTex via Latent Space and the Alexandr Wang statement to The Next Web; the edge scorecard exemption from the independent-score rule stated; the China draft export catalogue added to the licence watch; the Muse licence terms stated once in the architecture watch and pointed to elsewhere; glosses added for Terminal-Bench, the Intelligence Index, NVFP4, the KV cache, speculative drafters, the Fairwind Program, the 3-of-6 test and decision models.
• Scorecard: the closed-frontier challenger moves from GPT-6 Astra to Gemini 4 Argon on availability (tie at 53); the multimodal leader and challenger swap (Argon to leader, Gemini 3.8 Live Extended Thinking to challenger); the edge leader moves from EmbeddingGemma 2 to LFM2.5-24B-A2B and the challenger from the 4B decision heads (the four-billion-parameter decision models that held the slot in W41) to Nemotron 3.5 Bolt 74B-A9B 24GB, both on vendor figures under the exemption stated in the row. Open frontier, reasoning, coding and sub-dollar rows are unchanged.
• Evidence labels: 'outside the graded packet' marks independent figures gathered after the graded source set closed and not re-verified by the house. This week that covers the Artificial Analysis figures for Reflection Beam (index 30, Terminal-Bench 4.0 3.0%, 47% hallucination), Nemotron 3.5 Super VL (26), Muse Spark Open's agentic placement (Terminal-Bench 4.0 8.1%; 'last of the top nine' is Latent Space's reading of the tables), the MiMo / GLM / Kimi / Qwen open-tier scores, the Decision Index October 15-16 snapshot (d1-ultra 65.2, Jev 1.14 57.1), the Muse Spark Open and DeepSeek Flash cost-per-task rows, the NVIDIA-Reflection talks (FT), the China draft export catalogue, the Alexandr Wang statement (The Next Web) and the 'DOA' verdict (@teortaxesTex via Latent Space). Each is labelled at every use. Every vendor benchmark is labelled vendor-run or card at the point of use; DeepSeek's Terminal-Bench 4.0 is labelled as measured in the Claude Code harness wherever it appears.
• Fact homes: the week's open-weight licence terms and API rate cards live in this issue's architecture watch; OpenAI's ChatGPT Enterprise per-token card, Workday's Flex Credit meters and Microsoft's Copilot credit cap live in The Application Layer; the capital, hardware and networking context (RTX Spark, Intel GPU D-Series, the Anthropic financing) lives in the AI Stack Weekly; harness releases live in Agent Techniques. This issue carries the headline figures from each and points to the home for the rest.