Meta's Muse Spark is an open-weight model that runs on two Macs

Meta has released Muse Spark, which it calls the most capable US open-weight model to date. The weights are free to download under Meta's own license. It is a mixture-of-experts model with 816 billion total parameters but only 42 billion active at a time, a 1 million token context window, and native multimodal input. Meta aimed it at coding, long-horizon agent tasks, and multi-step reasoning.

The detail that got attention is where it can run. A pair of Apple M5 Ultra Mac Studios, each with up to 512 GB of unified memory, acts like one machine with about a terabyte of memory, enough to hold the model and run it locally through the MLX framework. Meta showed it making iOS code changes to an open-source Mastodon client and editing video in DaVinci Resolve, all on-device. Beyond Apple hardware it supports SGLang, vLLM, TensorRT-LLM, Hugging Face Transformers, llama.cpp, and Unsloth, with custom GPU kernels that Meta reports give kernel-level speedups between 1.03 and 3.95 times on NVIDIA and AMD. The license is the sticking point: it is the Muse Model License, with limited commercial terms rather than a true open-source grant, so the "open" here means open weights, not unrestricted use. Meta says it added extra safety post-training because of the risks of releasing weights, and that a full evaluation report is still to come.

Why it matters

Running a frontier-scale model on two desktop machines changes the calculus for teams that cannot or will not send code and data to a hosted API. Before you build on it, read the Muse Model License carefully, because the commercial limits, not the hardware, are what will decide whether you can ship with it.