Meta’s Muse Spark Open Is an Open-Weight Model With Strings Attached

Set Trending Topics as a preferred source on Google.

Meta is back in the open-weight game, with some fine print: The company has just released the weights of Muse Spark Open, which it calls “the most capable US open-weight model to date.” The model has 816 billion parameters, 42 billion of which are active per request, a context window of one million tokens and is built for coding and A.I. agents. According to Meta, it is the first model in its class that can even run on a desktop, using two linked Mac Studios with Apple’s M5 Ultra chip.

What the Model Can Do

Muse Spark Open is natively multimodal, meaning it understands images and video as well as text. In a demo, Meta had the model fix bugs in an iOS app on its own and review footage frame by frame in the video editing program DaVinci Resolve. The weights are available for download on Hugging Face, Amazon SageMaker, Google Cloud and Microsoft Foundry, and the model can be accessed via API through providers including Together AI, Fireworks, Groq and OpenRouter. Running it is not cheap, though: According to Unite.AI, the model needs at least 700 gigabytes of graphics memory depending on the format, and the two Mac Studios with 512 gigabytes each together cost about as much as a compact car.

For Meta, the move is a reversal. With Muse Spark, the company introduced its first model without open weights in the spring, breaking with its earlier Llama strategy. With Muse Wonder, Meta is now also competing at the very top among closed frontier models.

How It Compares

So far, the benchmarks come almost entirely from Meta itself; a detailed evaluation report is to follow. They show that Muse Spark Open clearly beats other American models such as Nvidia’s Nemotron 3 Ultra and Thinking Machines’ Inkling, often by more than 20 points. It does not, however, reach the level of the best Chinese open-weight models:

In Meta’s own comparison, GLM 5.3 (45), Kimi K3 (44) and Qwen3.8 (40) all rank ahead of Muse Spark Open on the Artificial Analysis Intelligence Index. Meta’s model leads, by contrast, on agent and tool tasks such as JobBench and MCP-Atlas. The comparison with Mistral is close: Mistral Large 4 scored 38 points on Artificial Analysis, while Muse Spark Open scored 39, according to the published data. That puts Meta and Mistral neck and neck for the title of best open model outside China.

The Artificial Analysis Ranking

Meta’s figures have not yet been independently confirmed: Muse Spark Open is not yet listed in Artificial Analysis’s ranking of open-weight models. For now, Chinese developers clearly dominate it:

Xiaomi’s MiMo-V2.6-Pro is the new leader. The highest-ranked American model is Thinking Machines’ Inkling in ninth place, and the only Meta model listed so far is the small Muse Glimmer, with 17 points. If Artificial Analysis confirms the 39 points Meta cites, Muse Spark Open would tie DeepSeek V4.1 Flash for fifth place and would be by far the best American model in the ranking. Mistral Large 4 is also missing from the list, because its weights have not yet been released and Artificial Analysis currently classifies it as proprietary.

What the Muse Model License Allows

The model is released under Meta’s own Muse Model License, which the company says has “limited commercial terms, which allow many developers to use the model for free.” It is a custom license with conditions that go beyond classic open-source licenses such as Apache 2.0 or MIT. According to the license text, the following is allowed:

• Use and distribution: The model may be used, copied and distributed free of charge.
• Modifications: Fine-tuning, derivative models and distillation are expressly permitted.
• Own products: Companies may use the model internally and as a component of their own products, including commercially.
• Outputs: Meta claims no rights to the results the model generates.

What the License Does Not Allow

The restrictions mainly affect large providers and certain areas of use:

• Model as a service above $20 million: Anyone offering the model as an A.I. service with more than $20 million (about €18 million) in revenue over 12 months must notify Meta in advance and sign a separate agreement. The threshold does not apply to internal use, research or building the model into one’s own products.
• Mandatory notice: Every product that uses the model or a derivative must prominently display “Built with Muse Spark Open,” for example in its interface, documentation or settings.
• Prohibited uses: The acceptable use policy bans, among other things, weapons development, malware, operating critical infrastructure, and fraud, disinformation and impersonation. According to the tech publication The New Stack, it also rules out military deployments.
• Age limit: The model is not intended for anyone under 18.

According to The New Stack, the license closely resembles the one under which Moonshot AI released its Chinese model Kimi K3. Other Chinese developers, such as Alibaba with Qwen and Zhipu with GLM, have so far mostly relied on more permissive licenses such as Apache 2.0 or MIT, which have neither revenue thresholds nor naming requirements. For most developers and start-ups, the Muse license is unlikely to pose much of a hurdle in practice. For large cloud and A.I. providers that want to resell the model, however, Meta remains the gatekeeper.