Meta's Muse Spark Open Is a Huge Open-Weight Model Built for Two Mac Studios
Meta released Muse Spark Open, an 816-billion-parameter open-weight model with a one-million-token window, and says it runs on two Apple M5 Ultra Mac Studios. Benchmarks and the full safety report are still missing.
Meta has released Muse Spark Open, an “open-weight” model, which means the trained model file is public so others can download and run it. It has 816 billion parameters in total, roughly the number of adjustable dials inside the model, and a context window of one million tokens, a rough measure of how much text it can hold in mind at once. Meta calls it “the most capable US open-weight model to date.”
The details
The model uses a “mixture of experts” design. Think of a large office where each question is routed to a few specialists instead of the whole staff. Only 42 billion parameters are active for any one answer, which keeps it faster than its size suggests. It handles text and images and is aimed at coding and long, multi-step agent work.
The eye-catching claim is the hardware: Meta says two Apple M5 Ultra Mac Studios, linked together and running through Apple’s MLX software, can host it, with up to a terabyte of combined memory. The weights are on Hugging Face, Amazon SageMaker, Google Cloud and Microsoft Foundry, and hosted versions exist at providers like Together AI and OpenRouter. Runtimes such as llama.cpp, vLLM and MLX are supported. The license is Meta’s own “Muse Model License” with limited commercial terms, so it is open in the download sense, not in the fully free-for-everyone sense.
What is behind this
Meta spent the summer shipping closed Muse models and a small open one called Muse Glimmer. This release returns to the open side with something far bigger. Open weights matter because anyone can run the model privately, inspect it and build on it, without sending data to a company’s servers.
A fair caveat: the announcement page has a “Key Benchmarks” heading but no actual numbers, and the full safety evaluation is only promised “soon.” Meta’s own demos are not independent tests.
What this means for you: For most people, nothing changes today. Two high-end Mac Studios cost a lot, so this is for labs, companies and serious hobbyists. If you are curious about local AI, the practical gain is indirect: big open models tend to get shrunk into smaller ones that do run on ordinary laptops within months. Wait for independent benchmarks before judging its quality.
Sources
Nvidia Opens Nemotron 3.5 Super VL and a Local Coding Model That Fits in 24 GB
Nvidia released the open Nemotron 3.5 Super VL, which reads documents and charts, plus Bolt, a compact model for local coding. The headline benchmark wins apply only to customised versions on narrow tasks.
