Model Releases
Muse Glimmer
Meta's Apache-2.0 30B Muse Glimmer targets local agentic tool use, multi-step coding, and vision, sized to leave headroom on 32GB machines.
Meta released Muse Glimmer, a 30B open-weight model under Apache 2.0, as a local-sized counterpart to the Muse Spark line. Simon Willison’s write-up is the readable primary account; the model card and GGUF builds are on Hugging Face.
What was released
- 30B weights, Apache 2.0 (not a Llama-style custom license)
- Aimed at end-to-end agentic tasks: scaffolds, tool schemas, multi-turn coding
- Vision-capable; Willison ran an 18.16 GB LM Studio build
- Claimed coverage includes DeepSearch QA, MCP-Atlas, τ-Bench, and SWE-Bench
Why it matters
A 30B Apache model that still fits beside other apps on a 32GB machine is the practical local-agent size, not a datacenter checkpoint. Willison’s Datasette trial used tool calls to answer how auth works, which is the workload this wiki’s local stack cares about.
"Meta are back in the open weights game."