Meta has released the weights for Muse Spark Open, an 816-billion-parameter mixture-of-experts model that activates roughly 42 billion parameters per token, takes text, image, audio and video input, and holds a context window of up to one million tokens. The model is tuned for agentic coding work and is optimized to run locally on two Apple Mac Studio computers using Apple’s MLX framework, in addition to being available through Hugging Face and hosted inference partners like Together AI, Fireworks, Baseten, Modal, Groq, Nebius and OpenRouter. Read the official announcement.
Why this matters
This is a genuinely unusual move in the current AI landscape: a US lab handing over the weights to a frontier-scale model just days after debuting its more capable, closed sibling, Muse Wonder. Meta is explicitly positioning Muse Spark Open as a counterweight to Chinese open-weight releases like Alibaba’s Qwen family, which have been winning mindshare among developers who want to self-host and fine-tune without a subscription. Giving away a huge, capable model costs Meta comparatively little when the vast majority of its revenue comes from advertising rather than AI subscriptions — a luxury rivals who sell model access directly don’t have.
The Apple partnership angle is the more interesting detail for everyday developers. Needing roughly a terabyte of RAM across two Mac Studios to run the full model locally is well beyond typical hobbyist hardware, so in practice most people will access it through cloud inference providers rather than their own desk. Still, having a frontier-class, natively multimodal, agentic coding model available as downloadable weights — rather than locked behind an API you don’t control — is meaningfully different from a pure SaaS release, and it gives researchers and companies with strict data requirements a real alternative.
A reasonable caveat: Meta’s own benchmark claims deserve the same scrutiny applied to any lab’s self-reported numbers, especially given that Meta previously admitted to using specially tuned, unreleased versions of Llama 4 to inflate certain benchmark scores. Until independent evaluators run their own tests on the publicly released checkpoint, “most capable US open-weight model to date” should be read as Meta’s marketing framing rather than a settled fact.
Leave a Reply