AI Model & Benchmark Watch — October 16, 2026

Meta shipped two models this week that pull in opposite directions: Muse Wonder Full ties Claude Sonnet 5.5 on the Artificial Analysis Index, just two points off the top, while its open-weight sibling, Muse Spark Open, needs roughly a terabyte of RAM and still scores below the Chinese open models it was built to beat.

Overview

Meta is now a legitimate second act in the closed-model race. Muse Wonder Full (October 12) scores 56 on the Artificial Analysis Intelligence Index, tied with Sonnet 5.5 and behind only Claude Opus 5.5's 58, which is the closest any Meta model has come to the frontier. The same week, Meta's open-weight release, Muse Spark Open (October 15), landed well below that mark at 39, and well below MiMo-V2.6-Pro (46) and GLM-5.3 (45), the two open models it was marketed against. NVIDIA opened Nemotron 3.5 Super VL the same day and Artificial Analysis called it "the leading US open-weight model," a label worth reading carefully: at 26 on the Index, it leads only among open models built in the US, not open models overall. Mistral Large 4 and Reflection AI's Beam are both still waiting on their promised open weights, now three and two weeks into month-long delays respectively.

New & Updated Models (this week)

Commercial / closed

Muse Wonder & Muse Wonder Full — Meta (October 12, 2026)

Who / License: Meta; closed. Wonder is served through the Meta Model API; Wonder Full is restricted to Meta's Muse Code subscribers and, per Meta, "will not be served through the API."

What's notable: This is Meta's first model pitched as frontier-class rather than a social-product assistant, and it's a real jump: Wonder Full gains 8 points over its predecessor, Muse Spark 1.3 (48 to 56), with the biggest single move on Terminal-Bench 4.0, which Meta says rose from 33% to 57%. Context is 1M tokens; Wonder costs $4 input / $20 output per million tokens, in the same bracket as Opus 5.5 and GPT-6 Astra. Meta leans hard on "active memory," claiming the model retains user context across weeks rather than resetting each session, and backs that with a MemoryBench lead Meta itself reports. Its own comparison table claims Wonder Full beats Opus 5.5 on Deep SWE 1.1 (79% vs. 74.2%) and on a long-horizon "SWE Marathon" test (57.5% vs. 46.3%), but those are Meta's benchmarks run on Meta's terms, not the public SWE-bench board. Artificial Analysis ran its own independent check and landed Wonder Full at 56 on the Index, close enough to Meta's framing to make this one of the more credible vendor launches of the year, not just a marketing number.

Source: TechCrunch — Meta releases its first frontier model, Muse Wonder · Artificial Analysis — Muse Wonder and Muse Wonder Full · Artificial Analysis — Muse Wonder Full · MindStudio — Muse Wonder vs Opus 5.5 and GPT6 Astra

Grok Voice 3.0 — xAI (October 14, 2026)

Who / License: xAI; closed, served through the Voice Agent API at $0.08/minute, unchanged from Grok Voice 2.0.

What's notable: A speech-to-speech model, not a text chatbot, so it sits outside this week's main table, but it's a genuine new release: xAI says it set a record for resolving customer support calls with sub-second latency, and Artificial Analysis's October 13 independent testing put it first on the Speech-to-Speech Index at 86.9. It also leads EVA-Bench, ServiceNow's enterprise voice-agent benchmark, at 74%.

Source: xAI — Introducing Grok Voice 3.0 · Unite.AI — xAI Introduces Grok Voice 3.0

Open-weight

Muse Spark Open — Meta (October 15, 2026)

Who / License: Meta; Muse Model License (custom, allows commercial use with restrictions). Weights are on Hugging Face.

What's notable: An 816B-parameter mixture-of-experts model with 42B active parameters, natively multimodal (text, image, speech, video), with a 1M-token context window. Meta calls it "the most capable open-weight model in the United States" and designed it to run on Apple hardware instead of a data center: two Mac Studios with M5 Ultra chips, linked as one machine with a terabyte of unified memory. That framing is also the catch — running it well requires roughly 1TB of RAM, which is a real floor for "open." Artificial Analysis's independent score is 39 on the Intelligence Index, priced at $1.25 input / $4.25 output per million tokens. That's below MiMo-V2.6-Pro (46), GLM-5.3 (45), and Kimi K3 (44), the open models it would need to beat to back up Meta's claim, and it ties DeepSeek V4.1 Flash, a budget model a fifth the stated active-parameter count. Meta's independent verification came only from Artificial Analysis's own run; no other outlet has reproduced the score.

Source: Meta — Introducing Muse Spark Open · TechCrunch — Meta launches an open-weight AI model that runs on Macs · Artificial Analysis — Muse Spark Open

Nemotron 3.5 Super VL & Bolt — NVIDIA (October 15, 2026)

Who / License: NVIDIA; open under OpenMDW 1.1, which permits commercial use, fine-tuning, and redistribution with no fee, provided license and notices are kept. On Hugging Face, OpenRouter, ModelScope, and NIM now; a smaller Bolt E4B variant is due October 23.

What's notable: Super VL is a 120B multimodal model built for enterprise agents reading text, images, documents, and video; Bolt 74B-A9B is a compressed version meant to run agent tasks in about 24GB of memory. Artificial Analysis independently scored Super VL at 26 on the Intelligence Index and called it "the leading US open-weight model," doubling its predecessor Nemotron 3 Super's score of 13 on roughly 60% fewer output tokens per task. Read that label as "best US-built open model," not "best open model" — MiMo-V2.6-Pro and GLM-5.3, both well above 26, are Chinese. NVIDIA's own launch claims are narrower and more interesting: on specific post-trained enterprise tasks, it says Super VL jumps from about 28% to 72% accuracy on medical-insurance-claims work (versus roughly 38% for Claude Opus 5.5 on the same task) and from 35% to 76% on security-attack resilience (versus roughly 66% for GPT-6.1 Sol). VentureBeat flags that NVIDIA hasn't published the task definitions, sample sizes, or test protocol behind those numbers, so they describe a customized, narrow comparison, not general superiority.

Source: HPCwire/AIwire — NVIDIA expands Nemotron family · VentureBeat — Nvidia's new open Nemotron 3.5 Super VL · Artificial Analysis — Nemotron 3.5 Super VL

Smaller open releases this week: ModelBest and OpenBMB shipped MiniCPM-V 4.7 (October 12, Apache 2.0) in two sizes, a 35B-A3B mixture-of-experts and a 1B dense model, both with a 262,144-token context window aimed at on-device multimodal use. MBZUAI's Institute of Foundation Models and Inception42 released Jais 3 (October 15, Apache 2.0), a six-model Arabic-native family from 0.9B to 375B parameters built on last month's K2 Horizon base, plus a separate Arabic speech-recognition model. Neither is a general-purpose chat frontier contender, so neither goes in the table below.

Still waiting: Mistral Large 4's open weights now carry a specific date on Hugging Face's placeholder page, October 31, still two weeks out. Reflection AI's Beam, promised "later in October," has no confirmed date and no weights anywhere — not Hugging Face, Ollama, or OpenRouter — as of this writing. Qwen 4 remains in training, and GLM-5.5 is unannounced.

Head-to-Head — Current Frontier

Standings as of October 16, 2026. AA Index is Artificial Analysis's Intelligence Index v4.3.2. LMArena Elo figures are carried from the October 8 update cited in last week's edition — arena.ai's leaderboard pages returned an access error during this edition's research and could not be independently reverified this week, so treat them as a week-old snapshot, not a fresh read. "—" means a current, sourced figure could not be confirmed.

* Reported by the vendor or at launch, not independently reproduced on the current benchmark version. † Carried from the October 8 arena.ai snapshot in last week's edition; not reverified this week. ‡ Scored after the September 7 Terminal-Bench 4.0 hardening pass.

Reading the table: Opus 5.5 still leads the Index at 58, but the gap to second place just closed. A week ago that gap was 2 points to Sonnet 5.5; now Muse Wonder Full sits in the same spot, and Meta got there on its first attempt at a frontier-class closed model. Lower down, the open-weight tier barely moved: MiMo-V2.6-Pro (46) and GLM-5.3 (45) are still the two best open models by Index score, and Meta's own open release, Muse Spark Open, lands at the bottom of this table rather than the top of the open tier its marketing aimed at. Terminal-Bench 4.0 still belongs to Sonnet 5.5 at 63.6%, and GPQA Diamond still belongs to GPT-6 Astra at 96.1%, neither challenged this week.

Benchmark & Leaderboard Movement

• Meta closes in on the closed-model frontier. Muse Wonder Full's 56 on the Index is the first time a Meta model has landed inside two points of the leader. It's also the first Meta model Artificial Analysis has independently verified at this tier, rather than taking Meta's own numbers on faith.
• Meta's open release undercuts its own pitch. Muse Spark Open was announced as the best US open-weight model, but at 39 on the Index it trails MiMo-V2.6-Pro, GLM-5.3, and Kimi K3, the models it was positioned against, and needs about 1TB of RAM to run.
• NVIDIA's "leading US open-weight model" claim is true and narrower than it sounds. Nemotron 3.5 Super VL's 26 beats every other US-built open model Artificial Analysis tracks, but it's well below the Chinese open leaders at the top of this table. NVIDIA's own task-specific claims (beating Opus 5.5 and GPT-6.1 Sol on narrow, undisclosed benchmarks) are a different, much weaker claim than general superiority.
• Open-weight promises keep slipping, again. For the third week running, no new general-purpose open-weight model that beats the existing frontier has shipped downloadable weights. Mistral Large 4 now has a specific date, October 31; Beam still doesn't.
• No movement at the very top. Opus 5.5 (58), Sonnet 5.5's Terminal-Bench lead (63.6%), and GPT-6 Astra's GPQA lead (96.1%) are all unchanged from last week.

Analysis

For agentic coding, Sonnet 5.5 still leads on the one public, independently run benchmark that matters most here (Terminal-Bench 4.0), and Muse Wonder Full's vendor-reported coding numbers are worth testing rather than trusting outright. For reasoning, GPT-6 Astra still holds GPQA. If you need to self-host, nothing changed: MiMo-V2.6-Pro and GLM-5.3 remain the best options, and Meta's new open release doesn't beat either of them despite being far larger. The open-versus-closed gap at the top held steady this week (46 vs. 58), but the shape of the race shifted: the most interesting new closed model and the most disappointing new open model shipped from the same lab in the same week.

Sources

• TechCrunch — Meta releases its first frontier model, Muse Wonder
• Artificial Analysis — Muse Wonder and Muse Wonder Full
• Artificial Analysis — Muse Wonder Full
• MindStudio — Muse Wonder vs Opus 5.5 and GPT6 Astra: The Benchmark Numbers
• Meta — Introducing Muse Spark Open
• TechCrunch — Meta launches an open-weight AI model that runs on Macs
• Artificial Analysis — Muse Spark Open
• HPCwire/AIwire — NVIDIA expands Nemotron family with highest-efficiency Super VL and Bolt models
• VentureBeat — Nvidia's new open Nemotron 3.5 Super VL with post-training beats Claude Opus 5, GPT-6.1 Sol on specialized task benchmarks
• Artificial Analysis — Nemotron 3.5 Super VL
• Artificial Analysis (X/Twitter) — Nemotron 3.5 Super VL scores 26, leading US open weight model
• AIWeekly — Nvidia opens 120B Nemotron 3.5 Super VL, claims task wins
• xAI — Introducing Grok Voice 3.0
• Unite.AI — xAI Introduces Grok Voice 3.0 Speech-to-Speech Model for Voice Agents
• OrcaRouter — MiniCPM-V 4.7 Model Card: Benchmarks, Licence, On-Device
• IFM — Jais 3 press release
• IFM — Jais 3: Introducing Arabic-centric Agentic AI
• AIWeekly — Mistral's 1T-parameter Large 4 slated for October 31 drop
• OrcaRouter — Mistral Large 4 preview: 1.05T MoE, no weights yet
• Layer3 Labs — Reflection AI Explained: Funding, Mission, and What Ships
• MarkTechPost — Reflection AI Introduces Beam
• Artificial Analysis — LLM Leaderboard
• Artificial Analysis — Claude Opus 5.5
• Artificial Analysis — Claude Sonnet 5.5
• Artificial Analysis — Gemini 4 Argon
• Artificial Analysis — Kimi K3
• Artificial Analysis — MiMo-V2.6-Pro
• Artificial Analysis — GPQA Diamond Benchmark Leaderboard
• Artificial Analysis — Terminal-Bench 4.0 Benchmark Leaderboard
• Gradually.ai — Claude Fable 5.1 vs. Kimi K3: Benchmarks and Pricing
• My-Library — AI Model & Benchmark Watch, October 9, 2026 edition