Meta's Muse Spark Open pushes free 816B model into cloud rivals' turf

Meta is offering a frontier-scale model that many developers can use free of charge, and it has placed it on all three major clouds on day one. That pressures paid closed-model APIs, especially in coding and agentic work.

Meta released Muse Spark Open on 15 October. The company calls it "the most capable US open-weight model to date", according to Meta AI Research on its company blog. The model is natively multimodal, with 816B total parameters, 42B active through a mixture-of-experts design, and a 1M-token context window. Meta says it targets long-horizon tasks, complex software engineering and multi-step reasoning.

Distribution is broad. Weights are on Hugging Face, Amazon SageMaker AI, Google Cloud Model Garden and Microsoft Foundry Models. Hosted APIs come from Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI. Supported engines include SGLang, vLLM, TensorRT-LLM, Transformers, MLX, llama.cpp and Unsloth. Meta says Helion kernels in SGLang deliver 1.03–3.95x kernel speedups on Nvidia and AMD GPUs.

The weights ship under the Muse Model License, which carries limited commercial terms. Meta says many developers can use the model free of charge.

Local deployment is also on the table. Meta says the model runs on two Apple M5 Ultra Mac Studios through MLX. Each machine has up to 512 GB of unified memory and 1.2 TB/s of bandwidth, and the pair acts like one machine with about 1 TB of memory.

On safety, Meta says the model received extra safety post-training and met the release thresholds of its Meta Superintelligence Scaling Framework. The full evaluation report is "available soon", the company said.

The strategy is reach. By putting weights on every major cloud and seven inference hosts, Meta lets customers run the model inside infrastructure they already pay for. Hosting providers gain a headline model to sell. Closed-model vendors lose some pricing power with buyers who can now test a large open alternative at no licence cost.

The caveats matter. The licence has limited commercial terms, so operators should read it before building a product on it. The blog gave no benchmark figures in the text we reviewed, and the "most capable" claim awaits independent testing. Watch for the promised evaluation report and for third-party results on coding tasks, where Meta is aiming.