Meta Releases Muse Spark Open: 816B Open-Weight MoE With 1M Context

MarkTechPost reports Meta released Muse Spark Open, an 816-billion-parameter open-weight mixture-of-experts model with 42 billion active parameters and a 1-million-token context window. The outlet says it leads open models on MCP-Atlas among cited evaluations, positioning the release as a major open multimodal and agentic coding model from Meta.

We’ll treat the MCP-Atlas lead as Meta-favored leaderboard context until independent labs reproduce it. The shape of the release is clearer: a huge sparse model aimed at long-context coding agents, shipped open-weight, with Apple and NVIDIA serving stories in the same news cycle.

That’s good for builders who want to own weights. It’s also another escalation in the open-vs-closed arms race, where “open” still means you need a small data center — or a partner API — to run the full thing. I’m watching whether MCP-Atlas dominance holds under third-party harnesses, and how fast competitors answer with their own open multimodal MoEs.

Companion coverage from TechCrunch, The Verge, and Hugging Face fills in Mac Studio requirements, MLX collaboration, and serving recipes.