Meta returns to open source: 816B new model, runnable on two Mac Studios

Meta releases open-weight model Muse Spark Open, with 816B parameters, 42B active parameters, 1M context, natively multimodal, aimed at code and agent tasks. The model is developed with Apple, optimized for Apple silicon unified memory architecture based on the MLX framework, and can run locally on two 512G M5 Ultra Mac Studios; the two machines cost over $40,000. Meta has adapted inference engines including SGLang, vLLM, TensorRT-LLM, and llama.cpp. Benchmarks show its coding and reasoning abilities are weaker than models such as Kimi K3, Qwen3.8 Flash Next, and GLM 5.3 Flash, but it leads on some agent metrics; comparison models come from different harnesses, and some scores are self-reported by vendors.