Meta and Apple put Muse Spark Open on Mac Studios
Meta released Muse Spark Open, an 816-billion-parameter open-weight model it calls its most capable U.S. open model yet, and worked with Apple so a quantized copy can run across two or four linked M5 Ultra Mac Studios using Apple's MLX framework. Zuckerberg and Apple's John Ternus publicly backed the demo; Meta AI chief Alexandr Wang told Axios the rivals had collaborated for months.
AI-assisted · US Brief desk · Sources listed below
What happened
Meta on Thursday released the weights for Muse Spark Open, a natively multimodal mixture-of-experts model with 816 billion total parameters, about 42 billion active per token, and a 1-million-token context window. Meta calls it the most capable U.S. open-weight model to date and says it is built for coding, agentic work and long-horizon reasoning. Weights are on Hugging Face and several cloud model gardens; APIs are live through partners including Together AI, Fireworks, Groq and OpenRouter. Meta says it worked with Apple so a quantized version can run locally across two M5 Ultra Mac Studios with 512 GB of unified memory each, or four with 256 GB each, using Apple's MLX framework and Thunderbolt RDMA links.
Why it matters
The release is a rare, public Apple-Meta engineering partnership and a high-profile proof that a frontier-class open model can run on linked desktop Macs instead of only in cloud GPU clusters. It also puts Meta's open-weight strategy beside Apple's pitch for private, on-device AI — without putting Meta inside Apple Intelligence.
More context
How outlets framed it: Axios cast the launch as a Silicon Valley détente after years of privacy fights between Apple and Meta. Meta's own post stressed open weights and desktop agentic AI. MacRumors and 9to5Mac led with the Mac Studio angle and noted that Apple and Meta publicly backed each other on the day. Meta AI chief Alexandr Wang told Axios the companies had been collaborating for months and joked that they were "excited for mommy and daddy to be playing nice again," crediting former Apple executive Daniel Gross with helping bring the sides together.
What Apple published: In a same-day Machine Learning Research post, Apple said Muse Spark Open needs roughly 493 GB in NVFP4 form for the weights alone, so it does not fit on one 512 GB Mac Studio. Apple reported decode speeds of about 25 tokens per second on two M5 Ultra Mac Studios with tensor parallelization and about 36.73 tokens per second on four, at an 8,192-token context — Apple's own benchmarks. Meta CEO Mark Zuckerberg thanked Apple hardware chief (and new Apple CEO) John Ternus on X; Ternus replied that it was "amazing to run a model of Muse Spark Open's caliber and scale this efficiently, right from your desk," The Verge and MacRumors reported.
Hardware and cost caveats: MacRumors noted that Apple has not yet released the 512 GB Mac Studio configuration highlighted in the demos and that the 256 GB model already lists above $10,000. Meta says the open weights ship under the Muse Model License with limited commercial terms. Axios and Meta both stress that local runs keep prompts on the user's machines, so Meta does not see that traffic — a point Wang said sidesteps long-running privacy disputes.
Timeline: In 2018, then-CEO Tim Cook publicly criticized Facebook over privacy; in 2021 Apple's App Tracking Transparency changes hit Meta's ad business and Zuckerberg reportedly told staff to "inflict pain" on Apple, Axios recalled. Meta first released Muse Spark as an open-weight line earlier in 2026, then Muse Wonder for paid agents on Oct. 12. On Oct. 15 came Muse Spark Open plus the joint Mac Studio/MLX demonstration.
What's still unknown: Independent benchmarks versus other open models; when (or whether) the 512 GB Mac Studio ships and at what price; whether Apple will ever use Meta models inside Apple Intelligence; and how widely developers will actually buy multi-Mac Studio clusters versus using the hosted APIs. Open-weight release also leaves the usual dual-use questions Meta itself flags in its safety note.
5 listed sources
- Meta AI Research, Oct. 15, 2026
- Apple Machine Learning Research, Oct. 15, 2026
- Axios, Oct. 15, 2026
- The Verge, Oct. 15, 2026
- MacRumors, Oct. 15, 2026
References listed by US Brief; a source count is not a verification score.
Editorial sourcing notes
Model specs, availability, license language and on-device demo claims come from Meta's Oct. 15 research blog. Apple's throughput figures and hardware requirements come from Apple's Oct. 15 MLX research post. Framing of the Apple-Meta détente and Wang's quotes come from Axios; Ternus and Zuckerberg posts follow The Verge and MacRumors accounts of their X posts. MacRumors' note that the 512GB Mac Studio has not shipped yet is attributed to that outlet. "Most capable US open-weight model" is Meta's claim, not an independent ranking.
Like US Brief? You can support it with a tip.