Muse Glimmer 30B
Meta's first open-weights model since Llama 4: 30B dense, Apache 2.0, logit-distilled from Muse Spark, built to run locally.
Reverses part of Meta's closed turn with a permissively licensed 30B agentic model that has a perception encoder and a speculative-decoding drafter. It was trained by distillation from Muse Spark plus RL and benchmarked against Gemma4-31B and Qwen3.6-27B on agentic, coding and multimodal suites.
- Date
- Monday, 10 August 2026
- Lab
- Meta
- Kind
- open-weights
- Access
- open weights
Figures
| Measure | Value | Measured by |
|---|---|---|
| Size / license | 30B parameters, Apache 2.0 weights on Hugging Face; runs on a single 24 GB GPU with quantisation per trade press | company |
Same day Meta said Muse Spark 1.2 would also be open-weighted 'soon'; as of The Register (2026-09-02) no date given and no Spark weights found.
Sources
- research.meta.ai/blog/introducing-muse-glimmer-open-agentic-model
- www.constellationr.com/insights/news/meta-releases-open-weight-muse-glimmer-model-open-mus
This record was checked against its sources on 6 October 2026. How we check