Meta shakes up the frontiers, open and closed

What it claims

Muse Spark Open (816B total / 42B active, 1M context) is the strongest American open-weight model by a wide margin, but not the strongest open model overall. The author places it:

• above GLM 5.2,
• near DeepSeek V4.1 Flash,
• behind GLM 5.3 and Kimi K3, which were released months earlier (Aug. 18 and Jul. 16).

The author judges it "good enough" for the agentic workloads that GLM 5.2-class models made practical. The overall ranking given is Anthropic > OpenAI > Meta ~ Gemini > xAI ~ Chinese open weights > U.S. open weights. Xiaomi V2.6 Pro is named as the leading open-weight model overall.

How it works

The page quotes the Hugging Face model card on what the model is:

• It is built for coding and agentic work.
• It is natively multimodal.
• It received additional safety post-training so it resists misuse when run standalone, without system guardrails.
• Meta recommends deploying it inside a broader system with use-case-specific guardrails.

Some observers suspected the safety training had "broken" the model, because it shipped after Muse Spark 1.3 yet scores lower. The author argues it is more likely a safety tune of Muse Spark 1.2, which would explain a drop of only about 1 point on the Artificial Analysis Intelligence Index. As support, the author cites Alex Wang saying at the Muse Glimmer launch (a 30B agentic model) that Muse Spark 1.2 would be released as open weights.

Evidence

• Mistral Large 4 scores 38 on the Artificial Analysis Intelligence Index, but its weights were not yet released at the time of writing.
• Muse Wonder / Muse Wonder Full (closed) trail only Anthropic's Opus 5.5 on traditional benchmarks. They are similar to GPT-6-Astra and Gemini-4-Argon. Real-world vibe checks are described as solid but not expected to match the refinement of the latest Anthropic and OpenAI models.

The page gives no per-benchmark numbers for Muse Spark Open itself; it only refers to a chart.

Limits

The licence terms that matter before adopting:

• Revenue gate: the $20M/month threshold counts revenue across your entire business, not just the product that uses Muse. Above it you need a separate licence from Meta, which the author ties to revenue-share deals for resold inference.
• Outputs and derivatives: the attribution requirement, the revenue gate and the acceptable use policy all extend to model outputs and derivatives. This makes it legally ambiguous whether you can use the model to generate synthetic training data.
• Internal use: internal research and internal applications are explicitly permitted.
• Untested: the licence has not been tested in court. The author expects Meta to pursue only egregious violations and notes that almost no one complied with Llama's equivalent attribution clause. That is the author's speculation, not a legal assurance.
• Open source: the acceptable use policy alone means the model does not qualify as open source.

The positioning claims are the author's assessments. Apart from the Artificial Analysis references, no evaluation methodology is given.