Meta released the weights for Muse Spark Open on October 15. It's a natively multimodal model with 816 billion total parameters, 42 billion of them active for each token, and a 1-million-token context window. Meta says it's built for coding and agent work and calls it "the most capable US open-weight model to date." The weights are on Hugging Face under Meta's own Muse Model License, which allows free commercial use with a few conditions, and the model is also available through several API providers.

What's in the model

According to the model card, Muse Spark Open is a mixture-of-experts model with 256 experts, 8 of which are used for each token. It takes text, images, video and audio as input and replies in text, and Meta says it was trained on data from more than 100 languages. It has four reasoning effort settings (low, medium, high and xhigh), with xhigh as the default.

How it scores

Meta's model card compares Muse Spark Open with other open-weight models. On the Artificial Analysis Intelligence Index, the card lists Muse Spark Open at 39, Qwen3.8 at 40, Kimi K3 at 44 and GLM 5.3 at 45, with Mistral Large 4 at 38. That's why Meta's wording says US: by the card's own numbers, three Chinese models score higher on that index.

On coding, the card's Terminal-Bench 2.1 results, all run with the same Terminus 2 agent, put Muse Spark Open at 76.4, compared with 82.0 for Qwen3.8, 83.9 for GLM 5.3 and 85.0 for Kimi K3. Its best showing in the table is on MCP-Atlas, a tool-use test, where its 89.2 is the top score among the models listed.

What the license allows

Meta describes the license as having "limited commercial terms" and calls the model open-weight rather than open source. The license lets you use, modify, fine-tune, distill, distribute and sell the model for free, but there are conditions. Any product that uses the model, or a model derived from it, must prominently display "Built with Muse Spark Open" in its interface or documentation. If you offer AI models to others through an API and your company's total revenue, including affiliates, is over $20 million in any 12-month period, you have to notify Meta and sign a separate agreement before offering this model that way. That rule doesn't apply to internal use, research, or using the model to power features in your own product. The attached usage policy says the model isn't meant for people under 18 and bans uses such as military applications and weapons development.

Where you can run it

Meta lists Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI as API providers. On OpenRouter, when we checked on the morning of October 19 Beijing time, five providers served it, all at $1.25 per million input tokens and $4.25 per million output tokens.

The weights can be downloaded from Hugging Face, Amazon SageMaker AI, Google Cloud Model Garden and Microsoft Foundry. On Hugging Face you need to request access, and requests are approved manually. The model card says running it yourself takes at least 700GB of GPU memory for the NVFP4 version and 2TB for the BF16 version. Meta also says it's the first model in its class that runs on a desktop setup of two M5 Ultra Mac Studios using Apple's MLX framework, since each machine can give the GPU up to 512GB of memory. Supported inference engines include SGLang, vLLM, TensorRT-LLM, Hugging Face Transformers, MLX, llama.cpp and Unsloth.