Meta's giant open model fits on your desk. Bring two Macs.
Muse Spark Open has 816 billion parameters, a million-token memory and an unusual co-star: Apple, which published the recipe for running it on Mac Studios the same day.
By The desk, News · · 2 min read
Tags: meta · apple · open-weights · models
Meta released the weights for Muse Spark Open on October 15. It is a mixture-of-experts model with 816 billion parameters in total, of which about 42 billion are active for each token, plus a one-million-token context window. Meta says it was built for coding and agent work and calls it the most capable US open-weight model to date. That ranking is Meta's own claim.
Developers can call it through Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI, or download the weights from Hugging Face, Amazon SageMaker, Google Cloud Model Garden and Microsoft Foundry. The licence is Meta's own Muse Model License, which Meta describes as having limited commercial terms. Open weights, in other words, not open source in the strict sense.
Apple did the plumbing
The surprise guest is Apple. Its machine learning team published a same-day post showing how to run the model locally with the latest version of its MLX framework. Squeezed to 4-bit precision, the weights take up about 493 gigabytes, which is too much for one machine. Apple's setup links two Mac Studios with M5 Ultra chips and 512 gigabytes of memory each, or four with 256 gigabytes, over Thunderbolt 5. It needs macOS 26.2 or later.
Apple reports about 25 tokens per second on two machines and about 37 on four, at an 8,192-token context. That is faster than most people read. Meta says it tested agent tasks entirely on device, including shipping fixes to an open-source iOS app.
Our read
Opinion: “On your desktop” is doing a lot of work here. Two top-spec Mac Studios on a desk is a serious hardware budget, not a hobby. But for a law firm, a hospital or a lab whose data is not allowed to leave the building, a frontier-class model that runs in a cupboard without a data center contract is a genuinely new option.
The bigger signal is political as much as technical. Chinese labs such as Alibaba's Qwen, Moonshot's Kimi and Z.ai's GLM have led open-weight models, as AFP noted. Meta wants that crown back. Apple, which has no frontier model of its own on this scale, gets to sell the box everyone runs it on. Two rivals, one shared enemy: other people's clouds.
What we could not check: Meta's benchmark charts and its safety summary. Meta says the model met the release thresholds of its own scaling framework and that a full evaluation report is coming soon. Until then, those are company claims.
Sources
The photograph is generated for this story. It does not show the people named in it.