Muse Glimmer 30B: An On-Device Agentic Model

Muse Glimmer 30B. An On-Device Agentic Model. By Arash Ebrahim Abtahizadeh and three hundred and eleven others, from Meta Superintelligence Labs. Published on the seventeenth of October. No code repository is linked to it, though it does publish its weights.

Muse Glimmer is an open-source causal language model of thirty billion parameters, paired with a perception encoder, and purpose-built for autonomous agentic tasks on consumer hardware. The word agentic is doing real work in that description. The model folds multi-step reasoning, reliable tool use and recovery from failure into a single system, rather than leaving them to separate components. And that single system runs locally, on a consumer device, without needing any cloud infrastructure behind it.

Text and images go in, and text comes out. For efficient inference on a single GPU, the weights are quantized to approximately four-bit precision, which is what lets a model of this size sit on ordinary hardware at all. On agentic benchmarks, the authors report state-of-the-art performance for its size class. The two comparison models are of similar size, Gemma4-31B and Qwen3.6-27B, and the comparison covers tool use, multi-step planning, and agentic task completion.

On where it comes from, the abstract gives one detail: the model was trained on data from more than a hundred languages.

It is released under the Apache 2.0 license. The stated aim is to foster open-source development of agents that are capable, safe and local, which makes it a release for anyone who wants an agent running on their own machine rather than through a remote service.

Muse Glimmer is defined less by any one skill than by where it is meant to live. Reasoning, tool use and failure recovery are packed into a model sized and quantized for one consumer GPU, and then handed out under a permissive license for others to build on.