Meta released the weights for Muse Spark Open on Oct. 15. The multimodal model has 816 billion parameters in total, and its model card estimates that local deployment needs 700GB of GPU memory with NVFP4 weights or 2TB with BF16 weights. Those requirements put it beyond ordinary desktop hardware, even though the weights are available to download.
Muse Spark Open accepts text, images, audio and video, and returns text. Meta describes it as a model for coding and other work that takes several steps, such as planning an action, using a software tool and responding to what the tool does. In its announcement, Meta says its Muse Code system used the model to make and build iOS changes to Ice Cubes, an open-source Mastodon client, and to operate DaVinci Resolve while reviewing footage. These are Meta's demonstrations, not independent tests.
Expert routing leaves the full weights to store
The model uses a mixture-of-experts design. Its card lists 816 billion parameters overall and 42 billion active for each token; eight of its 256 experts are selected in each expert layer. Routing limits which parts do work at a given step, but the full set of weights still has to fit in memory or be split across devices.
Meta provides BF16 and NVFP4 weight versions. Its model card estimates 2TB of GPU memory for BF16 and at least 700GB for NVFP4. Meta says its Apple Silicon demonstration runs across two Mac Studio systems with M5 Ultra chips, each with up to 512GB of unified memory. Those are unusually large local configurations. Developers without that hardware can use hosted APIs from Meta's listed partners, which is a different deployment route from downloading the weights and running them on a personal computer.
Meta reports a higher MCP-Atlas score and a lower Terminal-Bench score
Meta's model card reports an 89.2 score on MCP-Atlas, compared with 82.3 for Kimi K3 and 84.2 for GLM 5.3. On Terminal-Bench 2.1, however, Muse Spark Open scores 76.4, below Kimi K3 at 85.0 and GLM 5.3 at 83.9. The table lists different reasoning settings for the models, so these company-reported results are not a controlled independent comparison or a single overall ranking.
The weights come with a custom license
Meta released the weights under its Muse Model License. It allows copying, modification and commercial use subject to the license and usage policy. A product or service using the model must prominently display “Built with Muse Spark Open” in its interface or accompanying documentation. A company offering the model as an API service must contact Meta and sign a separate agreement if its aggregate gross revenue with affiliates exceeds $20 million in any consecutive 12-month period.
The model card also recommends adding system-level safety protections suited to each use case.
Photo: AzureSaturn via Wikimedia Commons, CC0 1.0; unmodified. The image shows an earlier Mac Studio, not Meta's two-M5-Ultra configuration. License.