Introducing Muse Spark Open, The Most Succesful US Open‑Weight Model to Date

For extra particulars about our evaluations, see the evaluation methodology report.

Optimized for Apple Silicon

Leveraging the most recent model of MLX, Muse Spark Open was optimized to run on Mac Studio with M5 Ultra, and collectively they’re one of the simplest ways to run a robust mannequin by yourself desktop.

Apple’s unified reminiscence structure lets the M5 Ultra’s built-in GPU use as much as 512 GB per Mac Studio, so two machines can maintain a whole mannequin. And the chip’s 1.2 TB/s of reminiscence bandwidth is quick sufficient to learn the small slice every token wants. The result’s two Mac Studios behave like one supercomputer with a terabyte of unified reminiscence.

We examined a spread of agentic workflows operating totally on-device, utilizing Muse Code powered by Muse Spark Open, together with delivery actual iOS adjustments to Ice Cubes (an open-source Mastodon shopper) — bug fixes, new options, full construct — and driving DaVinci Resolve finish to finish, with the mannequin reviewing footage body by body to seek out the strongest moments and crop them vertically.

Available Now

We’re sharing this work as a result of we would like everybody to have entry to the very best fashions. The open-weight model of Muse Spark ships with customized, auto-tunable Helion GPU kernels in SGLang, delivering 1.03-3.95× kernel-level speedups on NVIDIA and AMD GPUs, and a 3% end-to-end speedup on AMD MI350.

Safety and Preparedness

Releasing open-weight fashions is a part of how we convey the advantages of non-public superintelligence to everybody, and a part of how security work advances throughout the sphere.

Open weights additionally carry their very own considerations, and Muse Spark Open has extra security post-training for that cause. The outcomes display sturdy safeguards throughout danger classes measured.

We evaluated the mannequin consistent with our Meta Superintelligence Scaling Framework, overlaying probably the most extreme dangers, and it met the thresholds our framework requires for launch. Detailed outcomes are within the Muse Spark Open model card, with a full analysis report accessible quickly.

License

We’re releasing the weights below the Muse Model License with restricted business phrases, which permit many builders to make use of the mannequin totally free.

Get Started

You can customise Muse Spark Open in your use case by leveraging PyTorch’s TorchTitan to additional tune the mannequin. To run it on Apple silicon, use MLX, which pulls the weights straight from Hugging Face.

Muse Spark Open is obtainable now by Hugging Face: https://huggingface.co/collections/meta-models/muse-spark-open

MLX Adapter for operating Muse Spark Open on Mac Studio.

Hosted Inference Partners: Together AI, Fireworks, Baseten, Modal, Groq, Nebius, and OpenRouter.

Built or printed one thing with Muse Spark Open? Share with us on the Hugging Face Forums — we’d like to see what you create.