Meta's new open model runs on two Mac Studios. Read the license before you build on it.

What Meta released

On October 15, Meta released the weights for Muse Spark Open, a mixture-of-experts model with 816 billion total parameters, of which 42 billion are active for any given token. According to its model card it uses 256 experts, eight active per token, across 62 layers, and was trained on data in more than 100 languages.

Meta describes it as "the most capable US open-weight model to date." That is Meta's characterization. Alongside the model it released Muse Code, an AI coding terminal, and demonstrated the model fixing bugs in a Mastodon client and editing video in DaVinci Resolve.

The desktop claim

The headline feature is where it runs. Meta calls it "the first model in its class to run on your desktop using two Apple M5 Ultra Mac Studios." With up to 512 GB of memory available per machine, Meta says, "two Mac Studios behave like one supercomputer with a terabyte of unified memory."

Some early coverage described it running on a single Mac Studio. Meta's own model card and announcement both say two, so that is what we report.

On servers, the model card calls for at least 700 GB of GPU memory at NVFP4 precision, or 2 TB at full BF16. It supports the usual open-source serving tools, including vLLM, SGLang, llama.cpp and Unsloth.

Read the license

The weights ship under the Muse Model License, which Meta describes as having "limited commercial terms." For most developers it is permissive: you can use, modify, fine-tune, distill and sell what you build. Three conditions are worth knowing before you commit:

• Attribution. Products must "prominently display 'Built with Muse Spark Open'" in the interface or documentation.
• The $20M clause. A "Model as a Service" provider whose group revenue exceeds $20 million in any 12-month period "must notify Meta and enter into a separate written agreement." This does not apply to internal use, research, or using the model inside your own product, as long as that product is not itself a model service.
• A usage policy. Prohibited uses include military and warfare applications, malware, and operating critical infrastructure. The card also says the model "is not intended for individuals under the age of 18."

Downloading is gated: Meta asks for your full legal name, date of birth and organization.

How it compares — by Meta's numbers

On the Artificial Analysis Intelligence Index, Meta's table gives Muse Spark Open a 39, behind GLM 5.3 (45), Kimi K3 (44) and Qwen3.8 (40), and just ahead of Mistral Large 4 (38). To Meta's credit, its own table shows rivals ahead on several tests, including Terminal-Bench, where Kimi K3 scores 85.0 to Muse's 76.4.

These are Meta's figures, run by Meta. Independent results will say more.

What to watch

• Independent benchmark results, and whether they match Meta's table.
• Whether model-service companies sign Meta's separate agreement, or choose a model with a simpler license.