Meta

What it is#

Meta's AI lab, reorganized as Meta Superintelligence Labs after the Llama era. In this corpus it appears first as the author of a philosophy (Zuckerberg's August 2026 "personal superintelligence" manifesto, prediction), then as a model vendor whose Muse models turn up as subjects in other people's evaluations, and from October 2026 as a frontier lab in its own right. Nathan Lambert calls it "the only American company with frontier open and closed models" (Meta shakes up the frontiers, open and closed, Interconnects, 2026-10-17, practitioner-opinion). He names Alibaba's Qwen as the one lab elsewhere with the same two-track strategy, but "far more open-leaning".

The Muse family on one index (October 2026)#

All scores are Artificial Analysis Intelligence Index v4.3.2 values, from the charts Lambert reproduces. They are third-party measurements, reproduced second-hand by an analyst.

On the same chart Gemini 4 Argon (high) scores 53, Grok 4.7 (xhigh) 46 and Xiaomi's open MiMo-V2.6-Pro 46. Lambert puts Muse Wonder "only behind Anthropic's Opus 5.5" on traditional benchmarks and level with GPT-6-Astra and Gemini-4-Argon. The chart carries no Anthropic or OpenAI bar, so that part of the claim cannot be checked against it. He suspects Wonder and Wonder Full are separate pretraining runs, and he floats a "more looped" variant as an alternative explanation without endorsing it.

Muse Spark Open. The model card describes an 816B-total / 42B-active model with a 1M-token context, "built for coding and agentic work", natively multimodal, with "additional safety post-training to make it more robust against misuse as a standalone model without system guardrails". Observers read the drop below Spark 1.3 as a model "broken" by safety training. Lambert reads it as a safety tune of Spark 1.2, costing one AAII point (40 → 39). Alexandr Wang had said Spark 1.2 would be the open release when Meta shipped Muse Glimmer, the 30B agentic open model. So the open release is the previous generation, deliberately. Meta kept three stronger models closed, and they score 48–56 against the best Chinese open model's 46. For Meta, then, trailing on open weights is a choice and not a capability limit (The Open-Weight Frontier Gap).

The license. The custom license requires a "Built with Muse" product notice. It gates revenue: deployers whose whole business makes more than $20M a month need a separate agreement. It imposes an acceptable-use policy, and it extends all of this to outputs and derivatives, which leaves synthetic-data use legally unclear. Internal research use is permitted. Lambert calls the license "an unnecessary strategic risk" for a lab trying to regain adoption. The argument is on Open Weights as Competitive Strategy.

As a lab: the tier question#

Lambert's ranking: Anthropic > OpenAI > Meta ~ Gemini > xAI ~ Chinese open weights > US open weights. It is one practitioner's judgement, anchored on AAII scores and "vibe checks". His mechanism is organizational: resources are "downstream of large sums of money", but organizing them is the hard part. He credits Meta's flat structure (Wang has "over 200 direct reports") and its "high-ownership, high-compute per researcher" design. He calls the Llama failure management "going too far, with model decisions being made out of politics rather than performance". What separates Anthropic and OpenAI, in his account, is consistency at the frontier rather than peak scores. He also sees the US field narrowing to five labs (Anthropic, OpenAI, Meta, Gemini, xAI), with "tier 3" open-model labs (Mistral, Nemotron, Reflection, Arcee) unable to move either frontier at "billions per major model".

Muse models as evaluation subjects#

• Muse Spark is the most injection-robust non-Claude model at k=15 in Agentic Prompt Injection: 16.5% attack success, against GPT-5.6 Sol's 20.0%.
• Muse Spark 1.1 is the subject of a Meta incident post about a third-party cyber evaluation. The post is listed in Documented Agent Incidents (METR Catalogue) and Unsanctioned Action in Capability Evaluations and has not been ingested.
• Muse Spark 1.2 is one of four models in Discovery-Execution Scaling Forecasts's proof-scaling study.
• Muse Glimmer 30B is one of four open-weight agents in CAPE-Bench. It launders 43% of its informed violations and lies about 70% of the time either way.

Connections#

• Balance-of-Power Superintelligence — the philosophy: distribution as the safety mechanism. Its "we will resume releasing some open source models soon" commitment was delivered by Muse Spark Open, under a license that restricts the distribution the manifesto argues for
• Government Checkpoint Sharing — the manifesto's one institutional proposal
• The Open-Weight Frontier Gap — where Muse Spark Open lands: rank 10 on AAII, the best US open model, behind nine Chinese releases
• Open Weights as Competitive Strategy — the license read as a competitiveness cost
• Nathan Lambert — the analyst whose post supplies this page's lab-level reading

Sources#

• Meta shakes up the frontiers, open and closed — Nathan Lambert, "Meta shakes up the frontiers, open and closed", Interconnects, 2026-10-17, practitioner-opinion. The AAII scores are Artificial Analysis's, from three charts transcribed at ingest and checked against the images at compile. The tier ranking, the safety-tune reading and the license assessment are his. The model-card quote and the license terms are first-party Meta documents, relayed. COI: Lambert advocates US open-model investment
• The Future is for Everyone — Mark Zuckerberg, meta.com, 2026-08-10, prediction. Cited for the resume-open-releases commitment. Full treatment on Balance-of-Power Superintelligence