Open-weight models

Most of the models in these docs are hosted on Meta Model API — you call them over the API and Meta runs the infrastructure. Meta also publishes open-weight models that you download and run on your own hardware, so you own the weights, the runtime, and the data.

This section covers the open-weight releases. For the hosted API models and how to choose between them, see Models.

Hosted API or open weights

Hosted on Meta Model APIOpen-weight (self-hosted)
How you run itCall the API; Meta runs the serversDownload the weights and run the servers yourself
Where prompts and responses goMeta Model APIYour own infrastructure
What you needAn API keyA serving runtime (such as SGLang or vLLM) and your own GPUs or cloud compute
Where to startQuickstartThe Get the model page for the model below

You can also call open-weight models through third-party inference providers instead of running them yourself. Each model's page lists its hosted providers.

Available open-weight models

Muse Spark OpenMeta's 816B-parameter mixture-of-experts model with 42B active parameters, for agentic and coding work, with text, image, audio, and video input. Ships as BF16 and NVFP4 checkpoints.
Muse GlimmerMeta's 30B open-weight multimodal model, distilled from Muse Spark and built for local and edge deployment.

Next steps

Pick a model above and follow its Get the model page to download the weights and serve them, or see Models for the full lineup, including the models hosted on Meta Model API.