Meta released Muse Spark Open, an 816 billion-parameter model optimized for two high-end Mac Studios. Open weights broaden deployment options, but the hardware burden limits claims that frontier AI now runs on ordinary desktops.
Meta has released the weights for Muse Spark Open, bringing a frontier-scale coding and agent model outside its own data centers. The release expands developer choice, but the phrase "runs on a Mac" needs a large qualification: Meta's reference setup uses two Apple M5 Ultra Mac Studios with a combined 1 terabyte of unified memory.
Muse Spark Open is a sparse mixture-of-experts model with 816 billion total parameters and about 42 billion active for each token. It accepts text, images, audio and video, and supports a context window of roughly one million tokens. Meta built it for long-running agent tasks such as reading a codebase, planning changes, calling tools and correcting failed steps.
The company says the model can run through frameworks including Apple MLX, vLLM, SGLang and llama.cpp, with support across Apple, Nvidia, AMD and Intel hardware. Hosted API providers also offer access, so developers do not need to buy the reference machines to experiment with it.
Open weights change the deployment trade-off
Companies that self-host can keep prompts and outputs inside their own infrastructure, inspect the model more closely and avoid dependence on one API provider. The cost is substantial memory, operations work and responsibility for safety controls. Meta's documentation says serving still requires loading the full 816 billion weights even though only a fraction activate per token.
Apple's contribution is also strategically relevant. Its machine-learning team described optimizations for distributed inference over Thunderbolt 5 and the unified-memory architecture of M5 Ultra systems. The collaboration showcases Mac Studio as local AI infrastructure, though it does not make the model a mass-market desktop tool.
For Meta, the release strengthens its open-weight ecosystem and may encourage developers to build around Muse rather than closed rivals. The business payoff will depend on adoption, hosted usage and whether Meta can convert an open model into durable platform influence without absorbing all of the training cost itself.
Research and commentary are provided for information, not personalized investment advice. Verify material claims with the linked source and original company disclosures. Report a correction · About BTI
