Meta released Muse Spark Open on October 15, an open-weight model trained for agentic coding that also accepts images, audio and video as input. The weights are available through Hugging Face.
Specs
Per Hugging Face’s launch post, the mixture-of-experts model has about 816 billion parameters, with around 42 billion active per token, and a context window of up to one million tokens. Each MoE layer holds 256 experts and eight are selected per token. Output is text only. Checkpoints come in BF16 and NVFP4 formats; the BF16 version is about 1.5 TB, and the NVFP4 build is meant for NVIDIA Blackwell GPUs and can run on a single B200 node.
Meta describes the model as tuned on real developer workflows such as multi-file refactors and long debugging sessions. It is the full-size member of the Muse series, following the small dense Muse Glimmer that Meta released in August. Hugging Face calls it the most capable US open-weights model to date; that is the platform’s own characterization.
Availability and early rankings
Amazon says the model is now available in SageMaker JumpStart for deployment into a customer’s own AWS account. Artificial Analysis lists Muse Spark Open (xhigh) at 39 on its Intelligence Index, behind its top-ranked proprietary entries.
Why it matters
Along with Reflection’s Beam, it makes the second very large open-weight coding model in two days.