News & analysis
Meta Releases Muse Spark Open, an 816B Open-Weights AI Model
The open-weights version of Muse Spark can be downloaded free on four major clouds and used commercially under a limited license, putting a frontier-scale model in developers' hands.
Key points
- Meta released Muse Spark Open, the open-weights version of Muse Spark, under a license with limited commercial restrictions.
- The model has 816 billion parameters but activates roughly 42 billion per token, a sparse design that lowers the compute needed per token.
- Weights are downloadable on Amazon SageMaker AI, Google Cloud Model Garden, Hugging Face and Microsoft Foundry Models, with API access through Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI.
Meta released Muse Spark Open, the open-weights version of its Muse Spark model, making the weights downloadable for free under a license with limited commercial restrictions. The model is an 816-billion-parameter sparse mixture-of-experts system that activates roughly 42 billion parameters per token, according to Meta's blog.
The release puts a model of unusual size into free circulation. Muse Spark Open is a sparse mixture-of-experts design: rather than running all 816 billion parameters for every token, the model routes each token through a subset, activating about 42 billion parameters. That gap matters because compute cost scales with what is activated, not with the total stored in the weights, so a model this large can still be served at a fraction of the cost a dense model of the same size would require.
Distribution is broad from day one. The weights are available for download on Amazon SageMaker AI, Google Cloud Model Garden, Hugging Face and Microsoft Foundry Models. Developers who would rather not host the model themselves can reach it through API platforms including Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI. The license permits many developers to use the model for free, though it carries limited commercial restrictions.
Meta also positions the model for local, agentic work. According to Meta's blog, Muse Spark Open brings advanced agentic capabilities to desktop computers for the first time on Apple M5 Ultra Mac Studios. Agentic capability refers to a model's ability to carry out multi-step tasks on its own rather than only answering single prompts, and running that on a desktop machine shifts where such workloads can execute.
Read next
News & analysis
ECB Holds Rates at 2.50% but Keeps Another Hike on the Table
