If you'd told a tech reporter five years ago that Meta and Apple would team up on anything, they'd have checked whether you were feeling OK. Now Meta has released a frontier-scale open-weight model, and Apple's own engineers published the instructions for running it on a pair of Mac Studios. Hell is reportedly building a ski resort.
816 Billion Parameters, One Desk
On October 15, Meta released Muse Spark Open, an open-weight model built for coding and agentic work. It's a natively multimodal mixture-of-experts design with 816 billion total parameters, 42 billion active per token, and a one-million-token context window. Meta calls it "the most capable US open-weight model to date." That's Meta's own assessment, and independent reviewers say Chinese open models still lead in some areas.
The Mac angle is real, but it isn't cheap. Apple says the NVFP4 build needs about 493 GB just for weights, so it won't fit on a single Mac Studio. You need at least two M5 Ultra machines connected over Thunderbolt 5 with RDMA, running Apple's MLX framework. In Apple's own tests at an 8,192-token context, two Macs using tensor parallelism generated about 25 tokens per second, and four Macs reached roughly 37. Weights are on Hugging Face and the model runs in vLLM, SGLang, llama.cpp, Transformers, and others. Hosted providers such as Together AI and Groq also offer it.
"Local AI" Now Means a Shopping List
For businesses, the real story is control rather than bragging rights. A model this capable that runs on hardware you own means sensitive data (contracts, patient notes, source code) never has to leave the building. Plenty of compliance teams have been waiting for exactly that. It's not a cheap setup, but it's a one-time hardware cost, and hourly GPU rental for a model this size isn't cheap either.
Read the fine print, though. Meta says the release comes with "limited commercial terms" under the Muse Model License, so "open-weight" doesn't mean "do whatever you like." Read the license before you build a product on it. Meta is also keeping its stronger Muse Wonder models closed, so Spark Open is a very good opening offer rather than the best Meta has. The smart approach is to treat the model as a component you can swap out later. Build your integration so a newer or cheaper model can replace this one without rewriting everything.
Bottom line: frontier-class AI now fits on a desk, as long as it's a sturdy desk with a big budget.
If on-premises AI just became realistic for your organization, we can help you decide whether a local model, a hosted API, or a mix of both fits your data, budget, and risk tolerance before you buy any Mac Studios.
Source: Let's Data Science