Overview
- Meta published Muse Spark Open on Oct. 15 as an open‑weight multimodal model with 816 billion parameters, 42 billion activation parameters, and a 1‑million token context window designed for code, agent workflows and long‑range reasoning.
- The model’s weights and API access are available across multiple platforms and cloud marketplaces, including Hugging Face, Amazon SageMaker, Google Cloud Model Garden and several API providers, under a Muse Model License that restricts commercial use for most developers.
- Meta and Apple publicly showed Muse Spark Open running locally using Apple’s MLX framework on two M5 Ultra Mac Studios equipped with 512 GB unified memory each, a configuration that reduces the desktop hardware barrier but carries very high cost.
- Meta says the release received targeted post‑training safety alignment and emphasized that its top frontier models such as Muse Wonder will remain closed for safety and capability reasons, reflecting a deliberate open/closed release strategy.
- Independent analyses show Muse Spark Open performs especially well on agent‑style benchmarks but does not unambiguously top all global open models on programming and general reasoning tasks, and the model’s practical value will depend on integration into toolchains and cost of deployment.