Muse-Glimmer-30B-AZ-Focus
Meta's Muse Glimmer 30B, with the second-guessing turned down. About a quarter less thinking, the same accuracy, and no loss on agentic coding. Runs anywhere Muse Glimmer does.
It stops second-guessing itself
Muse Glimmer thinks efficiently on most questions. On the hard ones, some answers turn into long loops of"Wait…", "Hmm…", "Actually…", re-checking work it has already done, and those long answers are no more likely to be correct. AZ-Focus turns that habit down. It still reasons step by step and still checks its work. It just loops less, so the slowest answers come back much sooner.
Less thinking, same answers
Against Meta's release on the same questions, AZ-Focus cuts thinking on every suite at both of Muse Glimmer's thinking-heavy settings: by 22% to 33% at its default reasoning strength (high), and by 17% to 31% at xhigh, with accuracy level overall.
No loss as an agent
Thinking less can cost a model its careful checking when it works as an agent. Not here: on SWE-bench Verified, AZ-Focus resolved as many tasks as the original, and finished more of them within the step limit.
| SWE-bench Verified, first 50 tasks | Original | AZ-Focus |
|---|---|---|
| Tasks resolved | 29 of 50 | 32 of 50 |
| Finished within the 100-step limit | 37 of 50 | 41 of 50 |
mini-SWE-agent with a 100-step limit, default reasoning strength, one run per task, same llama.cpp setup for both. AZ-Focus also took slightly fewer steps per task: 67.4 on average, against 69.4. 27 tasks were resolved by both, 5 by AZ-Focus only and 2 by the original only: within noise on 50 tasks, so read this as no loss, not a gain.
A drop-in for Muse Glimmer
The same files, architecture, chat template and vision encoder as Meta's release, with some output-projection weights changed. Set the reasoning strength as usual (low, medium, high or xhigh) and use Meta's recommended sampling.
BF16
For transformers and vLLM, or as the source for your own quantizations. Hugging Face
GGUF
Q8_0 (29.6 GB) for llama.cpp, with Muse Glimmer's tool calls parsed natively, and the vision projector. Hugging Face
EXL3
4.0 bpw for ExLlamaV3, about 17 GB, for 24 GB GPUs: 19% less thinking on AIME 2025, 54 vs 56 of 60 correct against the same build of the original. Hugging Face
Weights only
No adapter, runtime patch or extra code. Only 54 output-projection matrices differ from Meta's release.
Limitations
- Small sets: 128 answers and 50 coding tasks. Differences of a few answers are within noise.
- Measured with thinking on, at reasoning strengths high and xhigh. Lower strengths, vision, other languages and very long contexts haven't been measured separately.
- Set an output limit: on rare hard problems the model can still think for a very long time.
- AZ-Focus is designed to change how long the model thinks, not what it will or won't do. We haven't re-run Meta's safety evaluations: follow Meta's guidance on deploying with guardrails, and its usage policy.
Apache 2.0, as for the original, and Meta's usage policy. Based on Muse Glimmer 30B by Meta.
Want to build on this?
We're an independent AI lab, looking for a few partners to build with at the frontier.