Meta releases Muse Glimmer, a 30B open-weight vision model optimized for agentic tasks

Meta has released Muse Glimmer, a 30-billion-parameter open-weight vision model licensed under Apache 2.0. The model is positioned for local use cases, with the author noting the 30B size fits comfortably on machines with 32 GB of RAM or more.

According to the announcement, Muse Glimmer is optimized for three capabilities: end-to-end agentic task completion, reliable tool use, and multi-step reasoning. The model’s claimed strengths include sustaining coherent plans across complex, extended workflows and invoking tools with precise schemas throughout those workflows.

The model’s performance is framed around success rates on full-task benchmarks such as DeepSearch QA, MCP-Atlas, τ-Bench, and SWE-Bench, which measure abilities like working within scaffolds, writing and debugging code, and resolving multi-turn requests from start to finish.

Independent testing by Simon Willison demonstrates the model in practice: he used an 18.16 GB version of Muse Glimmer via LM Studio to generate a pelican-themed artifact and to explore a codebase with a prompt about authentication in the Datasette project. The model’s outputs were produced using the llm-coding-agent plugin and the llm-lmstudio integration, with a patch applied for compatibility with LLM 0.32.

Willison also evaluated the model’s vision capabilities by asking it to describe a photograph of two brown pelicans on a rocky shoreline, producing a detailed natural-language description of the scene.