Muse Spark Open

Muse Spark Open is a model whose weights Meta released on October 15, 2026, so developers can download and run it themselves instead of only calling it through an API. Meta calls it the most capable US open-weight model to date and says it was built for coding and agentic work, is natively multimodal, and handles a context of up to 1M tokens (Meta). Hugging Face lists its inputs as text, images, audio and video, with text as the output. The weights come under the Muse Model License, which Meta describes as having limited commercial terms that let many developers use the model for free. The announcement does not spell those terms out, and its text gives no benchmark scores. Meta points to the model card for detailed results and says a full evaluation report is coming.

The two parameter numbers in its full name, 816B total and 42B active, explain how it can run outside a data center. It is a mixture-of-experts model: according to Hugging Face's technical walkthrough, each expert layer holds 256 experts, and a router picks 8 of them for each token, so only about 42B parameters do work at any one step. All 816B still have to be held in memory, but each token reads only a small slice. That is the basis for Meta's claim that the model runs on two Apple M5 Ultra Mac Studios using Apple's MLX framework. Each machine has up to 512 GB of unified memory, which gives about a terabyte together, and Meta says the chip's memory bandwidth is fast enough for the portion each token needs (9to5mac).

In data centers, the requirements are larger. Hugging Face's launch coverage lists the full-precision checkpoint at about 1.5 TB, run across four nodes of eight H200 GPUs, while a compressed NVFP4 version fits on a single B200 node.