Meta releases Muse Spark Open, an 816B open-weight MoE model with 1M context

Meta Superintelligence Labs has released Muse Spark Open, the open-weight version of its Muse Spark model, an 816B-parameter Mixture-of-Experts model that activates 42B parameters per token.

Meta says it is the most capable US open-weight model to date, leading on MCP-Atlas with 89.2 but trailing Kimi K3, GLM 5.3 and Qwen3.8 on Terminal-Bench 2.1.

The NVFP4 build can run on 2 Apple M5 Ultra Mac Studios via MLX, and hosted APIs are available from several providers.