Meta releases Muse Spark Open, an 816-billion-parameter open-weight model that can run on two Mac Studios
The model has a 1-million-token context window and is free for many developers under Meta’s new Muse Model License. Meta’s own benchmark table shows it ahead of rival open models on some tests and behind on most coding tests.
Illustration · not a photoHow this story was checked
- ResearchedWith AI assistance · 3 sources
- AI source checkEach sentence matched to its source
- Editor approvalMayank Sahu · 16 Oct, 12:59
- Published16 Oct, 12:59
Verify this story
- 3 sources, 2 official · 2 with an archived copy · see sources
- Every fact has a label and a source number · fact ledger
- What we searched, read and could not verify · research note
- Machine-readable: facts as JSON · text as Markdown
How this was made: researched, drafted and source-checked by AI (Claude, made by Anthropic), then published under the editor’s standing approval. A person did not read it line by line before it went live. AI policy
In short
Meta released Muse Spark Open on 15 October 2026, an open-weight mixture-of-experts model with 816 billion total and 42 billion active parameters and a 1-million-token context window. Weights are on Hugging Face under the Muse Model License, and Meta says the model runs on two Apple M5 Ultra Mac Studios.
What Meta released
Meta published Muse Spark Open on 15 October 2026 and called it “the most capable US open-weight model to date”.1 It is a natively multimodal mixture-of-experts model with 816 billion total parameters, 42 billion of them active for each token, and a 1-million-token context window.12 Meta says it is built for coding and agent work, including long, multi-step tasks.1
In September, Meta chief executive Mark Zuckerberg had said an open-weight version of Muse Spark would come “soon”.3
Where to get it
The weights are on Hugging Face, Amazon SageMaker AI, Google Cloud Model Garden and Microsoft Foundry Models.1 Hosted APIs are offered by Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI.1 It runs on inference engines including vLLM, SGLang, TensorRT-LLM, llama.cpp and Apple’s MLX.1
Running it on a desk
Meta says Muse Spark Open is “the first model in its class to run on your desktop”, on two Apple M5 Ultra Mac Studios connected through MLX.1 Each machine can give its GPU up to 512 GB of unified memory, so the pair acts like one system with about a terabyte.1
The licence
The Muse Model License lets most developers use the model for free.1 Companies that offer the model as a hosted service and earn more than $20 million in gross revenue over any 12 months must notify Meta and sign a separate agreement.2 Products must show “Built with Muse Spark Open”, and the usage policy bans uses such as malware, weapons and impersonation.2
How it compares
All scores below come from Meta’s model card.2
- Ahead of the open models Meta compared it with on JobBench (61.8), MCP-Atlas (89.2) and Humanity’s Last Exam without tools (26.1).
- Behind Moonshot’s Kimi K3 on coding tests such as Terminal-Bench 2.1 (76.4 against 85.0) and DeepSWE 1.1 (60 against 68).
- Behind Zhipu’s GLM 5.3 on the Artificial Analysis Intelligence Index (39 against 45).
No independent tests had been published when we checked.
Safety
Meta says the open version received extra safety training and met the release thresholds of its Superintelligence Scaling Framework. It says a full evaluation report will be “available soon”.1 Meta’s earlier closed models are covered in our story on Muse Wonder.
Fact ledger
How labels workTracked from this story
All trackersQuestions readers ask
Is Muse Spark Open free to use?
Yes for most developers. Companies running it as a hosted service with more than $20 million in gross revenue over 12 months need a separate agreement with Meta.
Can I run it on my own computer?
Meta says it runs on two Apple M5 Ultra Mac Studios with MLX. Most other setups need data-centre GPUs.
Is it better than Kimi K3 or GLM 5.3?
Meta’s own table shows it ahead on some tests and behind on most coding tests. No independent results were available.