Meta Releases Muse Spark Open, Teaming Up with Apple to Put an 816B-Parameter Model on Mac Studios

What happened

Meta announced Muse Spark Open on Thursday: 816 billion total parameters, about 42 billion active, a 1-million-token context window, natively multimodal, built for coding and long-horizon agentic tasks. Weights are live on Hugging Face, Amazon SageMaker AI, Google Cloud Model Garden and Microsoft Foundry; API access is available through Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI.

Alexandr Wang announced the release on X, writing that Muse Spark Open is "the most capable US open-weight model to date" and capping the thread off with "america!" Per tech media reports, Meta CEO Mark Zuckerberg wrote on X the same day: "Muse Spark Open — weights releasing today — runs great on new Mac Studios thanks to work with @johnternus and Apple," with a congratulatory reply from new Apple CEO John Ternus.

The collaboration works like this: Apple updated its MLX software toolkit so a quantized version can run across two M5 Ultra Mac Studios (512GB of unified memory each) or four (256GB each), linked with Thunderbolt RDMA to behave as one machine holding the entire model. Meta says it tested real agentic workflows entirely on-device, including fixing bugs, adding features and completing a full build for the open-source Mastodon client Ice Cubes, and editing video end-to-end in DaVinci Resolve. Per The Wall Street Journal, running it requires at least 1TB of memory, with hardware costs exceeding $40,000.

The two companies are longtime foes: they clashed publicly starting in 2018, and after Apple's 2021 privacy changes Meta said it lost more than $10 billion in ad revenue that year (per earlier Wall Street Journal reporting). Wang said in an interview: "We think that a lot of the world will run on open models, and they should be American."

Key facts

1. Scale and positioning: 816 billion total parameters, about 42 billion active, 1M-token context, natively multimodal; Meta calls it "the most capable US open-weight model to date," aimed at coding and agentic work. Meta-shared benchmarks show it leading open models including Nemotron 3 Ultra, Mistral Large 4, Qwen3.8, GLM 5.3 and Kimi K3 on agentic evaluations like MCP-Atlas and JobBench.
2. Acknowledged gap: Meta acknowledged that Chinese open models remain stronger on software development and some other tasks, including Alibaba's Qwen, DeepSeek and Moonshot's Kimi K3; closed Chinese models like Kimi K3 and GLM 5.3 still lead on Meta's cited Terminal-Bench 2.1 coding benchmark.
3. License: Weights are released under the Muse Model License: free for internal use, with companies earning over $20 million annually required to sign a separate agreement.
4. Sequencing: This is the second open-weight model Meta has released since overhauling its AI efforts last year; earlier this year, the closed flagship Muse Spark, the first product from Meta Superintelligence Labs, was its first release.