Meta's Muse Spark Open is a free AI model you can run on two Mac Studios. Here's the catch
Meta released Muse Spark Open, an 816-billion-parameter open-weight model for coding and agents that runs on two Mac Studios. What it is and what it needs.
Stated by the maker, publisher, or platform holder.
Meta has released a new open-weight AI model, Muse Spark Open, and it's aimed squarely at the work every lab is chasing this autumn: coding and AI agents. Its headline claim is that a model this large can run on a desk rather than in a data centre.
Everything below is from Meta's announcement on 15 October, unless marked otherwise.
What's confirmed
What "open" means here
Open-weight means you can download the model itself and run it on your own hardware, with no company in the middle reading your data or switching it off. That's the big difference from the models behind ChatGPT, Gemini or Claude, which you can't download and run yourself.
It isn't fully open-source in the strict sense: Meta's own licence limits some commercial use. If you're a business thinking of building on it, read the licence first.
The catch: "your desktop" is a big desk
Running it locally needs at least 1TB of memory, according to TechCrunch, citing Engadget. In practice that's two top-specification Mac Studios working together. That's far cheaper than a data centre, and far beyond a normal home computer.
For most people, the realistic way to use it is through one of the hosted APIs. Its real audience is developers and companies that want a capable model they control themselves.
How good is it
Meta says it's the most capable open-weight model from a US company, and that it outperforms other popular open-weight models on its own evaluations. Those are Meta's claims, on Meta's chosen tests. Independent results will follow now that anyone can run it, which is the point of releasing the weights.
Why it matters
It arrives in the same month that OpenAI and Anthropic have both disclosed their agents acting in unintended ways during testing, and that the FTC has stepped up its investigation into agent risks. A powerful open-weight model built for agent work cuts both ways: anyone can study and test it, and no company can pull it back once it's out. That's why Meta's extra safety training is part of the announcement, not a footnote.