Loading page…
Meta says the model runs entirely on two M5 Ultra Mac Studios. Its downloadable weights allow customization, but commercial use remains subject to license conditions.
Listen to this story
Meta’s October 15 release makes Muse Spark Open downloadable for coding and sustained, multi-step work, with local deployment as an alternative to hosted inference for teams equipped with two M5 Ultra Mac Studios. The model has 816 billion total parameters, 42 billion active parameters and a one-million-token context; Meta says the paired Macs can hold it in unified memory. Commercial use is subject to a limited Muse Model License, so open weights do not mean unrestricted use, and Meta’s full evaluation report has not yet been published.
The desktop configuration uses Apple’s MLX framework and up to 512 GB of unified memory per Mac; Meta cites 1.2 TB/s memory bandwidth.
Meta demonstrated Muse Code completing Ice Cubes bug fixes, feature work and a full build, as well as using DaVinci Resolve to select and crop footage.
Helion kernels in SGLang showed 1.03–3.95× kernel speedups on NVIDIA and AMD GPUs; end-to-end improvement was 3% on AMD MI350, not the Mac setup.
Meta released Muse Spark Open’s model weights on October 15, 2026, giving developers a downloadable model built for coding and multi-step work. The company says it can run entirely on two Apple M5 Ultra Mac Studios—a local alternative to hosted access—though the release comes with limited commercial terms.
Muse Spark Open has 816 billion total parameters, with 42 billion active. Its context window—the material it can work with in a request—holds up to one million tokens. Meta positions it for complex software engineering, multi-step reasoning and tasks that require sustained work rather than a single answer.
Meta optimized the model for M5 Ultra Mac Studios using the latest MLX framework. Each machine’s integrated GPU can use up to 512 GB of unified memory. Together, Meta says, the two machines can hold the entire model. The company also cites the chip’s 1.2 TB/s memory bandwidth as sufficient to read the portion needed for each token.
In on-device tests through Muse Code, Meta says the model completed bug fixes, new features and a full build for Ice Cubes, an open-source Mastodon client. Another demonstration involved operating DaVinci Resolve, reviewing footage frame by frame to select moments and crop them vertically.
Separate GPU optimizations ship as automatically tunable Helion kernels in SGLang, software used to run the model. Meta’s performance figures below apply to NVIDIA and AMD hardware, not the desktop configuration.
Meta reports gains for individual Helion kernels on NVIDIA and AMD GPUs.
Meta reports this whole-system improvement on AMD MI350.
Hosted access offers a route that does not require setting up the desktop configuration. Meta lists these distribution channels:
Developers can further tune the model using PyTorch’s TorchTitan. The weights use the Muse Model License with limited commercial terms, which Meta says allow many developers to use the model for free.
Meta says it added safety post-training for the open-weight release and that the model met release thresholds under its Meta Superintelligence Scaling Framework. The company points to its model card for detailed results, with a full evaluation report promised soon.
Loading discussion...
Join the conversation
Explain what would make local operation worthwhile—or not—for you.
Be the first to share a perspective or an experience.
Reader comments
Newest comments first. Replies stay oldest first.