Meta releases Muse Spark Open as an open model

On October 15, 2026, Meta released Muse Spark Open. The model can be downloaded from Hugging Face and run on your own servers or locally. Meta calls it the most capable US open model to date and says it is the first model in its class that can run on a desktop.

The downloadable version of Muse Spark

Muse Spark Open is the open version of Muse Spark from Meta Superintelligence Labs. According to the model card, it is a mixture-of-experts model with 816 billion parameters in total, 42 billion of which are active at each step. The context window is 1 million tokens. It reads text, images, video and audio and answers in text. It is built for coding, multi-step agentic work, tool use and reasoning over documents. The training data covers more than 100 languages; Meta notes that not every language was evaluated and that performance may drop outside the strongly supported set.

Where it runs and on what hardware

Besides Hugging Face, the model can be downloaded from Amazon SageMaker, Google Cloud Model Garden and Microsoft Foundry. For those who would rather not host it, it is available as an API through Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI. The hardware bar is high: about 2 TB of GPU memory at full precision and at least 700 GB in NVFP4. Meta's desktop suggestion is two M5 Ultra Mac Studios with 512 GB of unified memory each, used as one machine through Apple's MLX. SGLang, vLLM, TensorRT-LLM, Hugging Face Transformers, MLX, llama.cpp and Unsloth are supported, with PyTorch and TorchTitan suggested for fine-tuning.

Ahead on agent tests, behind on the general index

According to the table on the model card, Muse Spark Open scores 89.2 on MCP-Atlas, which measures tool use, and 61.8 on the JobBench work-task benchmark, the highest of the models compared. On the Artificial Analysis Intelligence Index it scores 39, behind Kimi K3 (44), GLM 5.3 (45) and Qwen3.8 (40). It scores 76.4 on Terminal-Bench 2.1 and 80.4 on OSWorld Verified, where Kimi K3 is higher. In Meta's demos the model fixes bugs in an iOS app and edits video frame by frame in DaVinci Resolve.

License terms

The model ships under Meta's Muse Model License, which is not among the OSI-approved open source licenses and carries limited commercial terms. Products that use it must show "Built with Muse Spark Open" in their interface, documentation or website. Companies that offer the model to others as a service and exceed 20 million dollars in revenue over any 12 consecutive months must notify Meta and sign a separate agreement. The usage policy bans uses such as weapons, malware, fraud and unlicensed legal, medical or financial advice, and the model is not intended for users under 18.

Meta says it added extra safety training for the open release and that the model met the release thresholds of its Superintelligence Scaling Framework, with a detailed evaluation report to follow. The refusal and honesty rates on the card come from Meta's internal tests. Meta recommends running the model inside a system with guardrails and adding human confirmation for irreversible agent actions. Organizations would do well to test it on their own workloads and watch for comparative results from independent researchers.