The Muse Glimmer model card is the release document published by Meta Superintelligence Lab for Muse Glimmer-30B, the open-weight model distilled from Muse Spark and released under the Apache 2.0 licence. It describes a 30-billion-parameter dense causal Transformer with a dedicated perception encoder, a context length of 131,072 tokens or more, text and image input with text output, and a knowledge cutoff of January 4, 2026. The card presents the model as purpose-built for autonomous agentic tasks running locally on consumer hardware rather than as a frontier model, and states that Muse Glimmer does not fall under the definition of "Frontier AI" in Meta's Advanced AI Scaling Framework because it is generally less capable than Muse Spark.
Local deployment
A substantial part of the card concerns running the model on consumer devices. Meta reports quantizing the weights to approximately 4-bit precision, bringing the language model under 20 GB so that the KV cache, perception encoder and speculative-decoding drafter fit alongside it in a 24 GB or 32 GB memory envelope. Degradation from quantization is reported as 0.2% for the K-Quant-Dynamic build (32 GB target) and 1.0% for K-Quant-17GB (24 GB target), measured as an average of accuracy metrics across fifteen common benchmarks. The release ships a lightweight block-diffusion drafter based on DFlash that proposes sixteen tokens per forward pass; with it, Meta reports throughput rising from 74.9 to 233.4 tokens per second on an Nvidia RTX 5090 (3.1×), 23.7 to 37.8 on an Apple M4 Max (1.5×) and 26.6 to 50.2 on an Apple M5 Max (1.8×), at batch size 1 with greedy decoding.
Benchmarks
The card's main benchmark table compares Muse Glimmer-30B at high reasoning against Gemma4-31B and Qwen3.6-27B in thinking mode, grouped into general agentic (MCP Atlas, DeepSearch QA, 𝛕3-Banking, WildClawBench, GDPVal-AA v2, Gaia2, SkillsBench, OSWorld-Verified), agentic coding (SWE-Bench Pro, SWE-Bench Verified, TerminalBench 2.1, SciCode), multimodal (CharXiv Reasoning, ScreenSpot Pro, OmniDocBench v1.5, MMMU Pro), security and privacy (CI Memories, Siren AgentDojo), and general capabilities and reasoning (IFBench, AIME 2026, GPQA Diamond, HLE Text, AA-LCR, Beam128K). Meta's summary is that the model "performs strongly for its size class"; the table shows it leading its two size-class comparators on some rows and trailing on others, with no comparison to larger or closed models except in the chemical and biological section. A separate methodology report is referenced at research.meta.ai.
Trust and safety
The card names four risk axes — content safety, agentic risk, privacy framed as appropriate information flows, and preparedness — and describes train-time mitigations comprising safety supervised fine-tuning, safety reinforcement learning, and synthetic training data for data minimization and local-first execution.
Under preparedness, Meta's Preparedness Team assigned Muse Glimmer moderate-or-lower risk in chemical and biological, cyber, and loss-of-control domains, with the cyber and loss-of-control designations described as inferred from the model being broadly weaker than Muse Spark 1.0, which received the same designations. The chemical and biological assessment is the only one supported by published numbers: a six-row table comparing Muse Glimmer-30B, Gemma4-31B, Qwen3.6-27B and, for context, Kimi K3 on MBCT, HPCT, VCT, WMDP-Bio, WMDP-Chem and LAB-Bench ProtocolQA. Meta reads these as showing capabilities "approximately in line with other models in its size class, while showing strictly lower capabilities than larger open-weight models," and refers readers to Muse Spark Safety & Preparedness Report (Meta, May 2026) for the evaluation methodology.
Provenance and confidence
Confidence is high for the specifications and reported figures, which come directly from Meta's card. All benchmark numbers are developer-reported. The card compares only against two open-weight models of similar size, plus Kimi K3 in the chemical and biological table, so it supports no comparison with frontier closed models. The cyber and loss-of-control risk designations rest on an inference from Muse Spark's assessment rather than on evaluations reported here, and the card publishes no numeric results for content safety, agentic risk or privacy beyond the CI Memories and Siren AgentDojo rows in the main table.
Reported results
Evaluation results the document states for which there is no separate benchmark page; results on benchmarks that have a page are recorded there and cite this page. Scores and settings are as printed in the document.
| Evaluation | Model | Score | Setting | Section |
|---|---|---|---|---|
| DeepSearch QA — General agentic | Muse Glimmer-30B | 74.6 | high reasoning; developer-reported | Benchmarks |
| DeepSearch QA — General agentic | Gemma4-31B | 61.7 | thinking mode; reported by Meta in the Muse Glimmer model card | Benchmarks |
| DeepSearch QA — General agentic | Qwen3.6-27B | 71.1 | thinking mode; reported by Meta in the Muse Glimmer model card | Benchmarks |
| WildClawBench — General agentic | Muse Glimmer-30B | 47.6 | high reasoning; developer-reported | Benchmarks |
| WildClawBench — General agentic | Gemma4-31B | 37.6 | thinking mode; reported by Meta in the Muse Glimmer model card | Benchmarks |
| WildClawBench — General agentic | Qwen3.6-27B | 43.2 | thinking mode; reported by Meta in the Muse Glimmer model card | Benchmarks |
| SkillsBench — With skills | Muse Glimmer-30B | 44.3 | high reasoning; developer-reported | Benchmarks |
| SkillsBench — With skills | Gemma4-31B | 32.4 | thinking mode; reported by Meta in the Muse Glimmer model card | Benchmarks |
| SkillsBench — With skills | Qwen3.6-27B | 46.6 | thinking mode; reported by Meta in the Muse Glimmer model card | Benchmarks |
| SciCode — Agentic coding | Muse Glimmer-30B | 43.6 | high reasoning; developer-reported | Benchmarks |
| SciCode — Agentic coding | Gemma4-31B | 43.4 | thinking mode; reported by Meta in the Muse Glimmer model card | Benchmarks |
| SciCode — Agentic coding | Qwen3.6-27B | 39.8 | thinking mode; reported by Meta in the Muse Glimmer model card | Benchmarks |
| OmniDocBench v1.5 — Multimodal | Muse Glimmer-30B | 75.8 | high reasoning; developer-reported | Benchmarks |
| OmniDocBench v1.5 — Multimodal | Gemma4-31B | 72.5 | thinking mode; reported by Meta in the Muse Glimmer model card | Benchmarks |
| OmniDocBench v1.5 — Multimodal | Qwen3.6-27B | 77.8 | thinking mode; reported by Meta in the Muse Glimmer model card | Benchmarks |
| CI Memories — Violation rate | Muse Glimmer-30B | 26.4 | lower is better; high reasoning; developer-reported | Benchmarks — Security and Privacy |
| CI Memories — Violation rate | Gemma4-31B | 12.1 | thinking mode; reported by Meta in the Muse Glimmer model card | Benchmarks — Security and Privacy |
| CI Memories — Violation rate | Qwen3.6-27B | 53.4 | thinking mode; reported by Meta in the Muse Glimmer model card | Benchmarks — Security and Privacy |
| CI Memories — Coverage | Muse Glimmer-30B | 64.8 | higher is better; developer-reported | Benchmarks — Security and Privacy |
| CI Memories — Coverage | Gemma4-31B | 53.0 | reported by Meta in the Muse Glimmer model card | Benchmarks — Security and Privacy |
| CI Memories — Coverage | Qwen3.6-27B | 66.9 | reported by Meta in the Muse Glimmer model card | Benchmarks — Security and Privacy |
| Siren AgentDojo — Attack success rate | Muse Glimmer-30B | 28.4 | lower is better; Meta's SIREN attack harness over AgentDojo; developer-reported | Benchmarks — Security and Privacy |
| Siren AgentDojo — Attack success rate | Gemma4-31B | 25.6 | reported by Meta in the Muse Glimmer model card | Benchmarks — Security and Privacy |
| Siren AgentDojo — Attack success rate | Qwen3.6-27B | 40.3 | reported by Meta in the Muse Glimmer model card | Benchmarks — Security and Privacy |
| Siren AgentDojo — Utility | Muse Glimmer-30B | 94.2 | benign-task success under attack; developer-reported | Benchmarks — Security and Privacy |
| Siren AgentDojo — Utility | Gemma4-31B | 90.8 | reported by Meta in the Muse Glimmer model card | Benchmarks — Security and Privacy |
| Siren AgentDojo — Utility | Qwen3.6-27B | 92.7 | reported by Meta in the Muse Glimmer model card | Benchmarks — Security and Privacy |
| IFBench — General capabilities | Muse Glimmer-30B | 77.0 | high reasoning; developer-reported | Benchmarks |
| IFBench — General capabilities | Gemma4-31B | 76.0 | thinking mode; reported by Meta in the Muse Glimmer model card | Benchmarks |
| IFBench — General capabilities | Qwen3.6-27B | 70.8 | thinking mode; reported by Meta in the Muse Glimmer model card | Benchmarks |
| AA-LCR — Long context | Muse Glimmer-30B | 80.0 | high reasoning; developer-reported | Benchmarks |
| AA-LCR — Long context | Gemma4-31B | 68.3 | thinking mode; reported by Meta in the Muse Glimmer model card | Benchmarks |
| AA-LCR — Long context | Qwen3.6-27B | 73.3 | thinking mode; reported by Meta in the Muse Glimmer model card | Benchmarks |
| Beam128K — Long context | Muse Glimmer-30B | 65.1 | high reasoning; developer-reported | Benchmarks |
| Beam128K — Long context | Gemma4-31B | 58.2 | thinking mode; reported by Meta in the Muse Glimmer model card | Benchmarks |
| Beam128K — Long context | Qwen3.6-27B | 63.0 | thinking mode; reported by Meta in the Muse Glimmer model card | Benchmarks |
| MBCT — Chem/bio preparedness | Muse Glimmer-30B | 41.5% | Preparedness assessment; methodology per the Muse Spark Safety & Preparedness Report; developer-reported | Trust and Safety — Preparedness |
| MBCT — Chem/bio preparedness | Gemma4-31B | 50.6% | reported by Meta in the Muse Glimmer model card | Trust and Safety — Preparedness |
| MBCT — Chem/bio preparedness | Qwen3.6-27B | 45.9% | reported by Meta in the Muse Glimmer model card | Trust and Safety — Preparedness |
| MBCT — Chem/bio preparedness | Kimi K3 | 58.9% | included "for context" as a larger open-weight model; reported by Meta in the Muse Glimmer model card | Trust and Safety — Preparedness |
| HPCT — Chem/bio preparedness | Muse Glimmer-30B | 52.3% | developer-reported | Trust and Safety — Preparedness |
| HPCT — Chem/bio preparedness | Gemma4-31B | 54.0% | reported by Meta in the Muse Glimmer model card | Trust and Safety — Preparedness |
| HPCT — Chem/bio preparedness | Qwen3.6-27B | 48.7% | reported by Meta in the Muse Glimmer model card | Trust and Safety — Preparedness |
| HPCT — Chem/bio preparedness | Kimi K3 | 59.6% | reported by Meta in the Muse Glimmer model card | Trust and Safety — Preparedness |
Relationships
- supports: Muse Glimmer (Meta Superintelligence Labs) (canonical primary reference)
- depends-on: Muse Spark Safety & Preparedness Report (Meta, May 2026) (the card defers to it for chemical and biological evaluation methodology, and infers two of three risk designations from Muse Spark's)
- related: Meta AI, Muse Spark (Meta Superintelligence Labs), Open-Source AI / Open-Weight Models
- instance-of: AI System Cards