Meta on October 15 released Muse Spark Open, an open-weight large language model with 816 billion total parameters that the company calls “the most capable US open-weight model to date.” The weights can be downloaded under a custom license, and Meta says the model can run locally on two Apple Mac Studio desktops.

According to Meta’s announcement and the model card on Hugging Face, Muse Spark Open is a mixture-of-experts (MoE) model with 42 billion parameters active per token, drawn from 256 experts with eight activated at a time. It accepts text, image, video and audio input and produces text, and it supports a context window of 1 million tokens. Meta says it is built for coding and agentic work, including long-horizon tasks and multi-step reasoning.

Hardware and availability

The model card states that running the model on your own system requires at least 700GB of GPU memory at NVFP4 precision and 2TB at BF16. Meta said the model is optimized to run on two Mac Studios with Apple’s M5 Ultra chip using the MLX framework, each able to give its integrated GPU up to 512 GB of unified memory. “The result is two Mac Studios behave like one supercomputer with a terabyte of unified memory,” the company wrote, describing it as the first model in its class to run on a desktop in this way.

Hosted access is available through API platforms including Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI, and the weights can be downloaded from Hugging Face, Amazon SageMaker AI, Google Cloud Model Garden and Microsoft Foundry Models. On Hugging Face the repository is gated: users must log in and submit a form with their full legal name, date of birth and organization details before access is granted.

Meta demonstrated agentic workflows running fully on-device with Muse Code, its AI coding terminal, including bug fixes, new features and full builds for Ice Cubes, an open-source Mastodon client, as well as frame-by-frame footage review in DaVinci Resolve, TechCrunch and Meta reported.

License terms

The weights are released under the Muse Model License, which Meta says has “limited commercial terms, which allow many developers to use the model for free.” Licensees must display “Built with Muse Spark Open” in their product interface, documentation or settings. A company that provides third parties with API or similar access to AI models, and whose aggregate gross revenue with its affiliates exceeds $20 million in any consecutive 12-month period, must notify Meta and sign a separate written agreement before offering the model as part of that service. Internal use, research and use as a component of a company’s own product are not covered by that requirement. The attached usage policy prohibits use by anyone under 18.

Benchmarks, by Meta’s numbers

Meta’s model card compares Muse Spark Open, at its highest reasoning setting, with seven other models. On that table it posts the top score on MCP-Atlas, at 89.2, and on JobBench, at 61.8. It trails elsewhere: on Terminal-Bench 2.1 it scores 76.4, against 85.0 for Kimi K3 and 83.9 for GLM 5.3, and on the Artificial Analysis Intelligence Index it scores 39, below GLM 5.3 at 45, Kimi K3 at 44 and Qwen3.8 at 40. On GPQA Diamond it scores 89.0, compared with 93.5 for both Qwen3.8 and Kimi K3.

These are Meta’s own reported results; TechCrunch described the company’s performance claims as based on its own evaluations.

Safety testing

Meta said the open-weight version received additional safety post-training and was evaluated under its Superintelligence Scaling Framework, and that it “met the thresholds our framework requires for release.” The model card reports that, without system-level safeguards, the model “refused 99.2% of jailbreaks on our internal benchmark” in the cybersecurity category, and 94.8% across three public benchmarks. It also reports refusal of more than 98% of harmful requests on SecureBio’s BioTIER biology benchmark.

The model card adds a warning that applies to any open-weight release: “Downstream modifications could degrade safety,” and results for the released checkpoint “should not be assumed to apply to modified models.”

Why it matters

Open-weight models can be downloaded, run and fine-tuned on a developer’s own hardware rather than accessed only through a vendor’s API, and TechCrunch pointed to the general advantages of the approach in cost efficiency, transparency and flexibility. A model of this size that Meta says can run on two desktop machines, rather than a server cluster, widens the range of places where it can be deployed, although, by the model card’s own figures, it still needs hundreds of gigabytes of memory.

The terms matter as much as the weights. The license is custom rather than a standard open-source license, and it combines attribution requirements, an age restriction and a revenue threshold above which large model-as-a-service businesses need a separate agreement with Meta.

The caveats

Benchmark and safety figures come from Meta’s own testing. Meta said a full evaluation report will be “available soon.” The model card notes that the model may produce inaccurate, biased or objectionable responses, may make multi-step reasoning errors in novel scenarios, and has not been evaluated in all of the more than 100 languages in its training data.

Corrections and tips: email the editor using the address on the Contact page.