Muse Spark Open arrived from Meta this week: a mixture-of-experts model with 816 billion parameters that the company bills as the strongest US open-weights release so far. A quantized NVFP4 variant and the main checkpoint are both available on Hugging Face.
It builds on Muse Glimmer, the smaller dense model that shipped in August. Because the two share a tokenizer and a chat template, teams running Glimmer can switch over without reworking their pipelines.
Big model, small compute bill
Every mixture-of-experts block holds 256 experts, and a router picks eight of them per token. Hence roughly 42B parameters do the work at inference time, even though memory must hold all 256 experts. In BF16 safetensors the checkpoint weighs about 1.5 TB; the 4-bit release gives up a negligible amount of accuracy for a much smaller footprint.
Images and video frames become tokens through a separate 1.8B perception encoder, which allows one model to read a screenshot, handle a voice note or follow a screen recording as it writes and fixes code.
Tuned for agentic work
Training leaned on genuine developer workflows — long debugging sessions, multi-file refactors, tool calls. Meta presents the result as a system that plans, acts, observes and self-corrects across extended tasks, with web browsing and images kept in the loop.
Given the political friction surrounding the subject, the open-weights framing stands out. Open models have turned into a proxy argument in both Washington and Brussels, with safety researchers and industry groups invoking them on opposing sides. By releasing an 816B checkpoint under a permissive license, Meta adds another large artifact to that fight.
