Muse Glimmer
Meta AI · · 5 days after Muse Spark 1.2
Verified
3 of 3 published values traced to a source2 primary sources, 2 archived
Why this status
Cited to the lab's own announcement and the official model card. Not verified because the context window could not be found in any cited primary source — the value may be correct, but this dataset cannot yet show where it came from. Modalities from the model card's own pipeline_tag "image-text-to-text" on meta-models/Muse-Glimmer-30B. Capabilities evidenced from the primary sources: agentic: “Muse Glimmer is a 30-billion-parameter model optimized for always-on local agent workflows.” | coding: “Muse Glimmer achieves strong success rates on full-task benchmarks including DeepSearch QA, MCP-Atlas, 𝛕-Bench and SWE-Bench, which measure its ability to work within sc” | reasoning: “Muse Glimmer supports different reasoning strengths to select the right balance between quality and speed.” | multilingual: “Muse Glimmer is trained on data from more than 100 languages.”.
What these states mean · How records are judged · Has this record changed?
A 30B open-weights agentic model distilled from Muse Spark, sized to run on a single consumer GPU: quantised to roughly 4-bit to fit under 20GB, and shipped with a DFlash speculative-decoding drafter. Meta's pitch is the always-on local agent rather than the frontier.
Details
| Released | August 10, 2026stated in research.meta.ai, huggingface.coMuse Glimmer was released on August 10, 2026.
|
|---|---|
| Company | Meta AI |
| Family | Muse |
| Era | Agentic systems |
| Type | Language model · LLM |
| Modalities | Text + Image → Text |
| Capabilities | Agentic, Tool use, Coding, Reasoning, Multilingual, Vision |
| Weights | Open weights |
| Licence | Apache-2.0 |
| Context window | 131K tokensstated in huggingface.coMuse Glimmer supports a context window of 131,072 tokens.
|
| Parameters | 30Bstated in research.meta.ai, huggingface.coMuse Glimmer has 30,000,000,000 parameters.
|
Benchmarks
Each row is a dated claim by a named party, not a property of the model. A later revision adds a row rather than replacing one. This project publishes no composite or overall score.
| Benchmark | Score | Reported by | Date |
|---|---|---|---|
| OSWorld-Verified | 65.9 | Vendor reported | August 10, 2026 |
| SWE-bench Pro | 51.2 | Vendor reported | August 10, 2026 |
| SWE-bench Verified | 76 | Vendor reported | August 10, 2026 |
| SciCode | 43.6 | Vendor reported | August 10, 2026 |
| ScreenSpot Pro | 75.4 | Vendor reported | August 10, 2026 |
| IFBench | 77 | Vendor reported | August 10, 2026 |
| AIME | 94.7 | Vendor reported | August 10, 2026 |
Current information, elsewhere
This record is a historical one: every figure above is traced to a dated primary source. For what is true today — live pricing, current rankings, availability — these are better maintained than anything here could be.
- Current price and providers OpenRouter ↗
- Current benchmark scores Artificial Analysis ↗
- Weights and model card Hugging Face ↗
Sources
Record status: verified · confidence 90/100 · 2 of 2 primary
Cited to the lab's own announcement and the official model card. Not verified because the context window could not be found in any cited primary source — the value may be correct, but this dataset cannot yet show where it came from. Modalities from the model card's own pipeline_tag "image-text-to-text" on meta-models/Muse-Glimmer-30B. Capabilities evidenced from the primary sources: agentic: “Muse Glimmer is a 30-billion-parameter model optimized for always-on local agent workflows.” | coding: “Muse Glimmer achieves strong success rates on full-task benchmarks including DeepSearch QA, MCP-Atlas, 𝛕-Bench and SWE-Bench, which measure its ability to work within sc” | reasoning: “Muse Glimmer supports different reasoning strengths to select the right balance between quality and speed.” | multilingual: “Muse Glimmer is trained on data from more than 100 languages.”.
- research.meta.ai Official announcement archived
- huggingface.co Model card archived
Primary means published by the organisation that made the model. Verified records require at least one.
Muse family
- Muse Spark April 8, 2026
- Muse Spark 1.1 July 9, 2026
- Muse Spark 1.2 August 5, 2026
- Muse Glimmer August 10, 2026
- Muse Spark 1.3 September 2, 2026
- Muse Wonder October 12, 2026
- Muse Wonder Full October 12, 2026
- Muse Spark Open October 15, 2026
Compare Muse Glimmer with another model →
See Muse Glimmer on the interactive timeline →