Meta releases Muse Spark Open's weights, tuned with Apple to run on two Mac Studios
Meta calls it the most capable US open-weight model, but its own model card scores three other open models higher on the first benchmark it lists.
Meta released the weights of Muse Spark Open on October 15, 2026. It's a mixture-of-experts model with 816 billion parameters in total, 42 billion of them active at a time, and a context window of a million tokens.
Mark Zuckerberg's own post led with where it runs. Meta worked with Apple on it, and the model card says it's optimized to run on two Mac Studios with M5 Ultra using MLX, Apple's machine learning framework.
Muse Spark Open -- weights releasing today -- runs great on new Mac Studios thanks to work with @johnternus and Apple.
What Meta says it is
Meta's AI account called it the most capable US open-weight model to date, built for coding and agentic work. It takes text, images, video and audio.
- Released
- October 15, 2026
- Parameters
- 816B total / 42B active
- Context window
- 1M tokens
- Runs on
- two Mac Studios with M5 Ultra using MLX
- License
- Muse Model License
The model card compares it with seven other open-weight models. On the Artificial Analysis Intelligence Index, the first row of Meta's own table, Muse Spark Open scores 39. GLM 5.3 scores 45, Kimi K3 44 and Qwen3.8-2.4T-A95B 40, so three of the seven score higher than Meta's model on the table it published itself.
It leads on other rows. On MCP-Atlas, a test of using tools, it scores 89.2, the best in the table. It trails on coding, though: 76.4 on Terminal-Bench 2.1 against Kimi K3's 85.0.
What the license lets you do
The Muse Model License reads like a permissive license, granting use "free of charge" with a catch for hosting companies. A firm that sells access to models through an API, and whose revenue tops $20 million over a year, has to notify Meta and sign a separate agreement before it offers Muse Spark Open that way. Running it in your own product doesn't count.
This open weights version has additional safety post-training to make it more robust against misuse as a standalone model without system guardrails.
Meta still recommends adding guardrails of your own. For agents that can take real-world actions, the card suggests things like human confirmation before anything irreversible.
What Apple said
John Ternus replied 44 minutes after Zuckerberg's post, and kept his attention on the desk the model runs on.
Congrats to the team! Amazing to run a model of Muse Spark Open's caliber and scale this efficiently, right from your desk. Mac Studio keeps pushing the boundaries of on-device AI!
Two Mac Studios is still a serious setup (the weights alone run to well over a terabyte at full precision). The model card also lists SGLang, vLLM, TensorRT-LLM and Llama.cpp among the frameworks it supports, for anyone running it on other hardware.
More on Meta
- Wyden and six senators ask Meta to justify research tax credits on its AI data centersOctober 15, 2026
- Meta releases Muse Wonder, the models Zuckerberg says bring it to the frontierOctober 12, 2026
- Meta, Sierra, Stripe and Walmart are writing a standard for how AI agents deal with shopsOctober 6, 2026
- Meta open-sources Muse Gadgets, so you can build hardware for its AI agentOctober 3, 2026