Meta has open-sourced Muse Spark Open, a new 816 billion parameter multimodal language model designed to assist developers with coding tasks. The mixture-of-experts model, released October 15, brings text, images, audio, and video into a single model optimized for agentic workflows—reading codebases, writing and debugging code, and calling tools. With approximately 42 billion parameters active per token and a one-million-token context window, it represents the most capable open-weight model available to U.S. developers to date. Meta trained Muse Spark Open on real developer workflows including multi-file refactors and long debugging sessions, enabling the model to plan, act, observe, and self-correct over extended tasks. The model's multimodal inputs allow it to read screenshots, process voice notes, and watch screen recordings while working on code, offering developers a unified interface for complex coding assistance. The mixture-of-experts architecture means only 42 billion parameters activate per token, allowing for faster inference while maintaining all 256 experts in memory. The release expands Meta's open-source AI capabilities for enterprise development following the earlier Muse Glimmer model. Available in BF16 and NVFP4 formats, Muse Spark Open can run across multiple GPU configurations, including NVIDIA Blackwell chips. Developers can fine-tune and deploy the model using frameworks like llama.cpp, vLLM, and SGLang, democratizing access to advanced multimodal capabilities.