Meta has released the weights for Muse Spark Open, a multimodal model built for coding and agentic tasks, with deployment options spanning Mac computers and data center hardware. The model has 816 billion total parameters, with 42 billion active per token, and a million-token context window.
The October 15 release accepts text, images, audio and video and generates text. Meta describes applications including software engineering, extended tool use and workflows that recover from failed actions. The downloadable model comes in BF16 and lower-precision NVFP4 versions.
Meta’s Apple deployment guide specifies two Mac Studio computers with M5 Ultra and 512GB of unified memory each, or four machines with 256GB each. MLX divides the model across the computers using direct memory transfers over Thunderbolt 5. The setup requires macOS 26.2 or later and a direct connection between every pair of Macs.
The server configurations include a four-GPU Nvidia Blackwell system for NVFP4 and eight GPUs for BF16. Meta also documents eight-GPU AMD Instinct configurations and an INT4 version running on Intel Xeon 6 processors. These are distinct hardware and precision combinations.
Meta says it used Muse Code with the model to make bug fixes and add features to the Ice Cubes Mastodon app. It also reports local DaVinci Resolve workflows that reviewed and cropped footage.
The weights are released under Meta’s Muse Model License. Users may modify, fine-tune, distill and redistribute the model, subject to its conditions and usage policy. Products using the model or derivatives must display the required Muse attribution in their interface or accompanying material.
If a licensee or an affiliate offers models as a service and their aggregate gross revenue exceeds $20 million in any consecutive 12 months, the licensee must notify Meta and sign a separate agreement before offering Muse or its derivatives through that business. The condition does not apply to internal use or to features of the licensee’s own product that are not themselves a model-as-a-service offering.
The release follows preview announcements from Mistral for Large 4 on October 6 and Reflection for Beam on October 5. Both announcements set out plans for weight downloads later in October. Meta’s model is already listed for download through Hugging Face.
Meta says the open-weight version received additional safety post-training and met its release criteria. Its model card recommends use-case-specific testing and system-level safeguards, including human confirmation for irreversible actions. Hosted access is also available through providers including Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI.
Sources: Meta AI, Meta, Meta Developers
–
By the Control Plane Editorial Team