Meta Releases Muse Glimmer, a 30B Open-Weight Agent Model That Runs on One GPU

Meta Superintelligence Labs released Muse Glimmer, a 30-billion-parameter open-weight model built specifically for autonomous agent tasks, licensed under Apache 2.0 and available on Hugging Face.

Unlike frontier models that require cloud-scale infrastructure, Muse Glimmer is designed to run entirely on a single consumer GPU: once quantized to 4-bit precision, it fits within roughly 20GB of memory, putting it within reach of a 24GB graphics card. Meta said the model decodes about 3.1 times faster than comparable open models and works with the vLLM inference framework out of the box.

The model includes a dedicated perception encoder, allowing it to interpret images, screenshots, charts, and documents alongside text, and was purpose-built for agentic workflows that plan, execute multi-step tasks, and recover from errors without constant supervision. The release underscores a broader industry push toward capable AI agents that can run locally, without a subscription or a round trip to the cloud.