Welcome Muse Spark Open: open-weight frontier multimodal mixture-of-experts model from Meta

Meta has released Muse Spark Open, an open-weight multimodal model that Hugging Face, in a post dated October 15, 2026, calls the most capable US open-weights model to date. It accepts text, images, audio and video, generates text, and offers a context window of up to one million tokens. Meta says it was trained for agentic coding on real developer workflows such as multi-file refactors and long debugging sessions.

The model has roughly 816 billion parameters but uses a mixture-of-experts design: each layer holds 256 experts, and a router picks 8 per token, so only about 42 billion parameters are active at a time. Deployment still needs memory for the full model, but Hugging Face says it runs at the speed of a 42B model. It is the full-size sibling of Muse Glimmer, a small dense model Meta released in August.

Weights come in a BF16 checkpoint of about 1.5 TB and a 4-bit NVFP4 version for Nvidia Blackwell GPUs. Support spans Transformers, llama.cpp, SGLang, vLLM and MLX, and reasoning cannot be switched off, only set as low as "low." Hugging Face reports a 2-bit quant reaching 18 tokens per second on a 512GB Mac M3 Ultra.

The company also fine-tuned LoRA adapters on 64 H100 GPUs in about two hours and built demos including a 3D asset generator and voice-driven browser agents. The release puts a very large open coding model within reach of well-equipped developers, though even the 3-bit version needs more than 256GB of RAM.