Apple and Meta team up to run 816B Muse Spark Open on Mac Studio
Meta and Apple buried the hatchet to optimize the massive 816B-parameter Muse Spark Open model for local execution on M5 Ultra Mac Studios.
Models · Source: MacRumors
What happened
Meta just dropped Muse Spark Open. It is an 816-billion parameter open-weight model with a one-million token context window. Meta calls it the most capable U.S. open-weight model to date. It is built natively multimodal for complex software engineering and multi-step reasoning. The model uses 42 billion active parameters during inference. It is available right now through API platforms like Baseten, Fireworks AI, Groq, and Together AI. You can also download the weights directly from Hugging Face or Google Cloud Model Garden.
The real shocker is the hardware partnership. Meta and Apple collaborated for months to optimize the model for Apple silicon using the MLX framework. You can run this massive model locally by chaining together two 512GB M5 Ultra Mac Studios. The machines pool their unified memory to act like a single desktop supercomputer with a terabyte of RAM. The integrated GPU architecture uses the chip's 1.2 TB/s of memory bandwidth to read the small slice each token needs.
Meta tested the model entirely on-device using Muse Code. The agent shipped real iOS bug fixes and features to Ice Cubes, an open-source Mastodon client. It also drove DaVinci Resolve end-to-end to review and crop video footage frame by frame. Former Apple executive Daniel Gross reportedly brokered the peace between the two tech giants. Meta AI chief Alexandr Wang joked to Axios that they are excited for mommy and daddy to be playing nice again.
Key facts
- 816 billion — Total parameters in Muse Spark Open
- 42 billion — Active parameters during inference
- 1 million — Token context window
- 1.2 TB/s — Memory bandwidth of the M5 Ultra chip
- 512 GB — Unified memory per Mac Studio needed to run the model
Why it matters
Local AI development just got a massive hardware upgrade path. Builders no longer need to rely entirely on cloud clusters to run frontier-level agentic models. If you have the budget for high-end Mac Studios, you can run complex, long-horizon coding tasks right on your desk with zero data leaving your machine. This changes the math for enterprise teams dealing with highly sensitive codebases. You pay a steep upfront cost for the hardware, but you eliminate recurring API fees and data privacy headaches.
Apple is positioning itself as the ultimate AI developer workstation. By optimizing MLX for massive open models, Apple creates a moat around its high-margin Mac Studio hardware. Meanwhile, Meta gets a premium, decentralized distribution channel for its open-weight models. This alliance sidelines traditional cloud providers and puts pressure on Nvidia's monopoly. It proves that Apple's unified memory architecture is a highly viable alternative for heavy AI inference. Competitors will have to rethink their local hardware strategies.
For builders
Local agentic coding for sensitive data
You can now run a 42-billion active parameter model locally for complex software engineering. Enterprise teams handling proprietary codebases win big by avoiding cloud API privacy risks. You keep your data completely on your own hardware.
High upfront hardware costs
Running this locally requires two 512GB M5 Ultra Mac Studios. With the 256GB version already priced over $10,000, founders will pay a massive premium. You trade high upfront capital expenditure for zero ongoing cloud inference fees.
Limited commercial use license
Muse Spark Open is free to download but comes with commercial restrictions. Builders need to read the fine print before baking this model into paid products. You might need to negotiate terms if your startup scales rapidly.
My take
I never thought I would see Apple and Meta hold hands. Mark Zuckerberg and John Ternus teaming up shows how AI is redrawing every alliance in Silicon Valley. If two Mac Studios can replace a cloud server for agentic coding, the local AI revolution is finally here. I am watching this space closely because hardware monopolies are finally breaking.
Original reporting: MacRumors. This is my rewrite and opinion.