Introducing Muse Glimmer

Meta is back in open weights with Muse Glimmer, a 30B model released under Apache 2.0. That alone will get attention. The old “Llama” licensing mess is apparently not the mood here anymore, and the new model is being pitched much more cleanly to people who actually run models on their own machines.

The big promise is not just text generation. Meta says Muse Glimmer is tuned for end-to-end agentic task completion, which in practice means it is meant to stay on task through full workflows instead of just sounding clever for a few turns. The company points to strong results on DeepSearch QA, MCP-Atlas, 𝛕-Bench, and SWE-Bench, the sort of benchmarks that check whether a model can work inside scaffolds, write and debug code, and handle multi-turn requests from start to finish.

There’s also the tool-use angle. Meta says the model can make precise function calls through extended workflows, and it claims Glimmer can sustain longer chains of reasoning without losing the thread. That combination matters more than another flashy demo prompt. A local model is only useful if it can keep doing the unglamorous parts: call the right tool, remember the goal, and not wander off.

Simon Willison’s own tests lean into that practical angle. He ran it in LM Studio using an 18.16 GB version of the model, tried it with his llm-coding-agent plugin on a fresh checkout of Datasette, and used it to answer “how does auth work?” He also noted that this size of model leaves room for other apps on a machine with 32 GB of RAM or more, which is exactly the kind of boring constraint that decides whether a local model is fun or just theoretical.

And Glimmer is not only a text model. Willison also asked it to describe an image, and it gave a detailed read of pelicans on a rocky shoreline, plus several smaller birds scattered around the frame. So Meta isn’t shipping this as just another chat toy. It wants a local model that can reason, use tools, and look at images without immediately falling apart.