Meta released the weights of Muse Spark Open on October 15. It is a natively multimodal mixture-of-experts model with 816 billion parameters, 42 billion active per token, and a 1 million token context window. Weights are on Hugging Face, Amazon SageMaker AI, Google Cloud Model Garden and Microsoft Foundry Models, and hosted APIs run through Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI. Hugging Face lists about 1.6 TB of weight files and gates download behind an acceptance form. The license is Meta's own Muse Model License, listed as "other" on Hugging Face. It allows use, modification, fine-tuning and distillation, and requires the notice "Built with Muse Spark Open". A licensee whose aggregate gross revenue with affiliates exceeds $20 million during any consecutive twelve-month period must notify Meta and sign a separate agreement before offering the model or derivatives as model-as-a-service via an API or similar interface. In Meta's scorecard, Spark Open (xhigh effort) scores 61.8 on JobBench against 61.4 for GLM 5.3, and 89.2 on MCP-Atlas against 84.5 for Qwen3.8. It trails Kimi K3 on Terminal-Bench 2.1 (76.4 against 85.0) and GLM 5.3 on GDPval-AA v2 (1401 against 1653). For a team that wants control over where a model runs, a large model with downloadable weights is available now, but the storage and memory bill is large and most teams will start on a hosted endpoint. The "Built with" notice and the $20 million clause are the license terms to check first. The column lineup and effort levels are Meta's: xhigh for Spark Open, max for GLM 5.3 and Kimi K3. For Beam and Mistral Large 4 it cites their self-reports. Treat the table as a reason to run your own task, since Meta says third-party runs are not tuned to each rival. Run your own task on a hosted Spark Open endpoint before sizing any hardware. Meta's scorecard for Muse Spark Open, October 15. JobBench 61.8 (GLM 5.3 61.4), MCP-Atlas 89.2 (Qwen3.8 84.5), Terminal-Bench 2.1 76.4 (Kimi K3 85.0), GDPval-AA v2 1401 (GLM 5.3 1653). Efforts differ by column: xhigh for Spark Open, max for GLM 5.3 and Kimi K3. Beam and Mistral Large 4 are self-reported, and Meta's runs of other models are best effort. Credit: Meta. Source: Introducing Muse Spark Open, The Most Capable US Open-Weight Model to Date, 2026-10-15. Muse Spark Open evaluation methodology: Lists each model's effort and sampling settings, which results are Meta's own runs versus leaderboards or self-reports, and the harnesses behind the coding rows. Muse Model License and usage policy: The attribution requirement, the $20 million model-as-a-service clause, derivative-model terms, and the usage policy (not intended for people under 18; bars unlicensed professional practice).