Meta releases Muse Spark Open, an 816-billion-parameter model built for coding agents

Meta has released the weights for Muse Spark Open, a large multimodal model aimed at coding and agentic work. The model has 816 billion parameters in total, with 42 billion active, and a 1-million-token context window. Meta says it is designed for long-running tasks such as complex software engineering and multi-step reasoning.

The weights are available on Hugging Face and other cloud model platforms, while hosted inference is offered through partners including Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI. The release uses the Muse Model License, which Meta describes as permitting many developers to use the model for free but includes limited commercial terms.

Meta says the model is natively multimodal, handling image, audio, video and document inputs, and is trained to plan, use tools and correct its work across longer tasks. PyTorch’s release notes describe capabilities including web browsing, computer use across desktop, mobile and browser environments, and coordinating subagents. Those are descriptions from the companies, not independent evaluations of how well the model performs in practice.

Meta also says it tested the model on two Apple M5 Ultra Mac Studios, using Apple’s MLX framework, and ran workflows including making changes to an open-source iOS app and editing video in DaVinci Resolve. The company says the two systems’ unified memory can hold the full model; this is a demanding desktop setup, rather than evidence that the model will run on an ordinary personal computer.

The release includes Helion GPU kernels and integrations with inference engines including SGLang and vLLM. Meta and PyTorch report kernel-level speedups of 1.03 to 3.95 times on NVIDIA and AMD hardware, and a 3% end-to-end speedup on AMD MI350. These are vendor-reported results, and the sources do not provide independent benchmarking in the material supplied.

Meta says it added safety post-training for the open-weight release and that the model met the company’s release thresholds under its safety framework. The company says a detailed evaluation report is forthcoming; the model card is cited as a source for further results. Open weights allow developers to run and adapt a model, but also mean it can be used outside the safeguards of a hosted service.