Meta releases Muse Spark Open, an 816B-parameter open-weight multimodal MoE for coding agents with a 1M-token context

Meta’s Muse Spark Open is an open-weight mixture-of-experts model with ~816B total parameters, ~42B active per token, and a context window up to 1,048,576 tokens. It accepts text, images, audio, and video but generates text, and is trained for agentic coding tasks such as codebase reading, multi-file refactors, debugging, and tool use. The article lists BF16 and NVFP4 checkpoints; in Transformers, BF16 runs across 4 nodes of 8 H200 GPUs, while NVFP4 targets a single B200 node.

For builders, the reported advantage is a single interface for long-context coding workflows that can also inspect screenshots, audio, or video. The trade-off is deployment cost: all 816B parameters must stay in memory, the BF16 checkpoint is about 1.5 TB, and even the NVFP4 path is described for NVIDIA Blackwell B200 hardware; the article does not provide independent benchmark results or clarify performance on less specialized infrastructure.