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

What happened: Meta released Muse Spark Open, an open-weight multimodal mixture-of-experts model in the Muse series, following Muse Glimmer. It is trained for agentic coding and accepts text, images, audio, and video. Key numbers: ~816B total parameters, ~42B active per token; 256 experts per MoE layer, 8 selected per token; up to 1M token context; BF16 checkpoint about 1.5 TB; NVFP4 checkpoint for NVIDIA Blackwell GPUs. Why it matters: Meta calls it the most capable US open-weights model to date. All experts stay in memory, so deployment needs 816B-class memory but runs at 42B speed.

Meta released Muse Spark Open, an open-weight multimodal mixture-of-experts model in the Muse series, following Muse Glimmer. It is trained for agentic coding and accepts text, images, audio, and video.

Meta calls it the most capable US open-weights model to date. All experts stay in memory, so deployment needs 816B-class memory but runs at 42B speed.