Meta responds to Qwen, GLM, and Kimi in the open-weight track with Muse Spark Open

Meta released the open-weight model Muse Spark Open, with 816B parameters (42B active), supporting up to 1 million token context and text, image, video, and audio inputs. Meta says it is the strongest open-weight model in the US, aimed at coding and agent tasks, leading Qwen 3.8, GLM 5.3, and Kimi K3 on several benchmarks such as JobBench, MCP-Atlas, HLE-Diamond, and ZeroBench, but trailing on some coding benchmarks and still clearly behind Meta's paid Muse Wonder Full model. The model is optimized for Apple M5 Ultra and can run locally on two Mac Studios with 512GB unified memory each; weights are available on Hugging Face.