Meta released Muse Spark Open on October 15, giving developers access to an open-weight version of its Muse Spark model for coding and other multistep agent tasks. Teams can run and adapt the model themselves or use a hosted inference provider instead of relying solely on Meta's managed products.
The model accepts text, images, video and audio and produces text. Its mixture-of-experts architecture contains 816 billion parameters, with 42 billion active per token, and supports a one-million-token context window. The smaller active count reduces the computation used for each token; it does not mean the full model occupies the memory of a 42-billion-parameter model.
What running it locally involves
Meta lists BF16 and NVFP4 checkpoints and launch support across inference tools including vLLM, SGLang, MLX and llama.cpp. The model card specifies at least 700GB of GPU memory for NVFP4 and 2TB for BF16. Those requirements describe the named checkpoints, not every third-party quantization.
For Apple hardware, Meta's deployment guide describes two M5 Ultra Mac Studios with 512GB of unified memory each, or four with 256GB each. Both configurations provide 1TB in total. The machines communicate through Thunderbolt 5 and RDMA, while MLX splits the model across them. The same guide lists a more compressed Unsloth version at about 300GB of combined memory. These are different deployment configurations, not equivalent quality or performance guarantees.
Developers who do not want to provision that hardware can use third-party providers. At publication, OpenRouter lists $1.25 per million input tokens and $4.25 per million output tokens for its Muse Spark Open listing. That is a hosted-service price, not a fee for downloading the weights, and other providers can set different rates.
Open weights with conditions
The Muse Model License permits use and modification subject to its terms, including a requirement to display “Built with Muse Spark Open.” Model-as-a-service providers whose aggregate company-and-affiliate revenue exceeds $20 million in any consecutive 12 months must notify Meta and obtain a separate agreement. Internal research and non-model-as-a-service uses are excluded; this is not a blanket revenue cap on all users.
Meta says this release adds safety post-training for standalone deployment. Teams still need to evaluate its behavior on their own tasks and apply appropriate safeguards.