Welcome Muse Spark Open: open-weight frontier multimodal mixture-of-experts model from Meta
Which summary reads better? Pick one — models revealed after.Both summaries are AI-generated.
Meta dropped an 816B-param MoE (42B active/token, 1M context) open-weight model natively ingesting text, image, audio, and video, tuned specifically for agentic coding—multi-file refactors, long debug loops, tool-calling, and web browsing with screenshots. It already has day-one support in vLLM, SGLang, Transformers, MLX, and llama.cpp plus an NVFP4 quantized checkpoint, so you can self-host a frontier coding agent without API dependency—but budget for the serving footprint: even at 42B active, 816B total weights demand serious multi-GPU memory even quantized.
Muse Spark Open is an 816B-parameter open-weight multimodal MoE from Meta with about 42B active parameters per token and up to a 1M-token context window. For production agent builders, the practical shift is that one open model can handle long-codebase coding workflows plus screenshots, audio, video, and tool use through a text-generation interface, but serving it will require frontier-scale MoE infrastructure rather than ordinary local deployment.