Meta's 816-billion-parameter Muse Spark Open model runs locally on linked Mac Studios
Representative image: tstokes / Pixabay
Meta released Muse Spark Open, an 816B-parameter open-weight model, on October 15, 2026, built with Apple to run on two to four linked M5 Ultra Mac Studios.
What Happened
Meta released Muse Spark Open, an open-weight artificial intelligence model with 816 billion parameters, on Thursday, October 15, 2026, and worked with Apple to make it run locally on Apple's new Mac Studio computers with the M5 Ultra chip. Meta describes the model as natively multimodal, with 42 billion parameters active at a time and a context window of one million tokens, aimed at coding and agent-style tasks. Because the model is too large for a single machine, Meta's recommended set-up links two M5 Ultra Mac Studios with 512GB of memory each, or four with 256GB each, using Apple's open-source MLX software. Meta chief executive Mark Zuckerberg said the model runs well on the new Mac Studios thanks to work with Apple's John Ternus.
Key Facts
- Model: Muse Spark Open, released by Meta on October 15, 2026, with downloadable (open) weights.
- Size: 816 billion total parameters, 42 billion active per step; context window of one million tokens.
- Hardware: two M5 Ultra Mac Studios with 512GB of unified memory each, or four with 256GB each. A compressed 4-bit version alone takes roughly 493GB, so it cannot fit on one machine.
- The 256GB Mac Studio is available now; the 512GB option is expected to ship in late October, according to reports.
- A Mac Studio with the M5 Ultra and 256GB of memory starts at $10,799.
- Software: Apple's MLX framework, which spreads the work across Macs linked by Thunderbolt 5 and uses remote direct memory access (RDMA) to move data between them quickly.
- Apple published its own benchmarks for two-Mac and four-Mac set-ups.
Why It Matters
Most powerful AI models today run in the cloud, in data centres owned by a few companies. Users send their questions over the internet and pay per use. A model that runs on machines in your own office changes that trade-off. Data stays on premises, there is no per-query bill, and the system can keep working without an internet link. For hospitals, law firms, banks or government offices with strict rules on where data can go, local AI can be the only kind they are allowed to use.
The release is also a meeting of two rivals with different strengths. Meta has built its AI strategy on open-weight models that anyone can download, while Apple's Mac Studio offers large pools of unified memory that suit local AI work. Linking several Mac Studios into a small cluster to run an 816-billion-parameter model shows how far that approach can stretch. It is still expensive: a set-up of four 256GB machines costs well over $40,000 at the starting price.
The model's design helps. Only 42 billion of its 816 billion parameters are active at any moment, a technique that lets a very large model run with less computing power per answer. The one-million-token context window means it can read very long documents or large codebases in one go, which is why Meta is aiming it at coding and agents that carry out multi-step tasks.
| Item | Detail |
|---|---|
| Total parameters | 816 billion |
| Active parameters | 42 billion |
| Context window | 1 million tokens |
| Two-Mac set-up | 2 x M5 Ultra, 512GB each |
| Four-Mac set-up | 4 x M5 Ultra, 256GB each |
Impact
Short-term: Developers and companies with the budget can download the model and run it on linked Mac Studios, testing whether an open model on local hardware can replace some cloud AI services. Demand for high-memory Mac Studios may rise, especially once the 512GB version ships.
Long-term: If local clusters prove practical, more organisations may choose to keep sensitive AI work in-house. The release also adds pressure on other model makers to publish open weights, and on chip and PC makers to offer machines with very large memory.
Who is affected: Software developers, companies in regulated fields that need data to stay on their own premises, researchers who want to study large models directly, rival AI labs, and cloud providers whose business depends on renting out AI computing.
Key Takeaway
Meta's 816-billion-parameter Muse Spark Open, released on October 15, 2026, can run on two to four linked Apple Mac Studios, bringing a frontier-scale open model out of the data centre and onto office hardware.
Questions and Answers
What is Meta's Muse Spark Open?
It is an open-weight AI model released by Meta on October 15, 2026, with 816 billion parameters, 42 billion active at a time and a one-million-token context window, designed for coding and agent tasks.
Can Muse Spark Open run on a single Mac?
No. Even a compressed 4-bit version needs about 493GB of memory. Meta's set-up links two M5 Ultra Mac Studios with 512GB each, or four with 256GB each.
Why does running AI locally matter?
Local AI keeps data on the user's own machines, avoids per-query cloud charges and works without an internet connection, which matters for organisations with strict data rules.
How much does the hardware cost?
A Mac Studio with the M5 Ultra chip and 256GB of memory starts at $10,799, so a four-machine set-up costs more than $40,000 at starting prices.
Disclaimer: Prepared by the Peepals newsroom from publicly available sources with AI assistance. Information is accurate to the best of our knowledge at the time of publication and may change. Images may be representative. Not professional advice. Report an error via our contact page.
Sourced and fact-checked by the Peepals Global Editorial Team
Amazon's Bee launches $74.99 Frontier and $99.99 Eclipse always-listening AI devices
Construction robotics start-up Origin raises $20 million led by Accel, Zacua
Reported, fact-checked and published by the Peepals Global Editorial Team.
