1. Summary

Muse Glimmer-30B is a dense causal Transformer with approximately 29.6B parameters, released by Meta Superintelligence Lab in August 2026. It is a multimodal model distillation from Muse Spark, featuring a dedicated perception encoder.

It is designed for the purpose of completing autonomous agent tasks on consumer hardware. In benchmarks, it demonstrates performance exceeding the comparison models Gemma4-31B and Qwen3.6-27B in areas such as MCP Atlas (75.5), DeepSearch QA (74.6), and SWE-Bench Pro (51.2). It shows strong performance in agent-based tasks.

On 2026-08-28, NVIDIA released Muse-Glimmer-30B-NVFP4, which was quantization using Model Optimizer.

2. Bunrin Bench (BUNRIN Lab Original Test)

No data (scheduled for implementation). For the methodology, refer to About Bunrin Bench.

3. Various Benchmarks

The following figures are listed in the model card (Hugging Face). Comparison targets are Gemma4-31B (Thinking Mode) and Qwen3.6-27B (Thinking Mode). Bold values indicate the highest score in that category.

Category Benchmark Muse Glimmer-30B (High Reasoning) Gemma4-31B Qwen3.6-27B
General Agentic MCP Atlas (Public) 75.5 54.2 62.5
DeepSearch QA 74.6 61.7 71.1
τ3-Banking 23.5 15.1 16.7
WildClawBench 47.6 37.6 43.2
GDPVal-AA v2 953 811 1141
Gaia2 43.3 36.4 40.0
SkillsBench (with skills) 44.3 32.4 46.6
OSWorld-Verified 65.9 58.5 75.6
Agentic Coding SWE-Bench Pro 51.2 36.9 50.2
SWE-Bench Verified 76.0 66.6 77.2
TerminalBench 2.1 51.7 43.4 60.7
SciCode 43.6 43.4 39.8
Multimodal Charxiv Reasoning 78.8 77.7 78.4
ScreenSpot Pro 75.4 75.9 76.1
OmniDocBench v1.5 75.8 72.5 77.8
MMMU Pro 74 73 75
Security & Privacy CI Memories (Violation ↓ / Coverage) 26.4 / 64.8 12.1 / 53.0 53.4 / 66.9
Siren AgentDojo (ASR ↓ / Utility) 28.4 / 94.2 25.6 / 90.8 40.3 / 92.7
General & Reasoning IFBench 77.0 76.0 70.8
AIME 2026 94.7 89.2 94.1
GPQA Diamond (AA) 83.5 85.7 84.2
HLE Text (AA) 22.0 23.6 23.1
AA-LCR 80.0 68.3 73.3
Beam128K 65.1 58.2 63.0

Source: Benchmarks section of the Hugging Face model card "Muse Glimmer Model Card."

4. Official announcements

  • 2026-08-28 NVIDIA released Muse-Glimmer-30B-NVFP4. This is a quantization model using Model Optimizer.
  • 2026-08-11 Ollama v0.32.8: Added Muse Glimmer support for NVIDIA, AMD, and other platforms. It became possible to launch from all platforms with a single ollama run muse-glimmer command.
  • 2026-08-11 Ollama v0.32.7: Provided initial Muse Glimmer support via the MLX engine for Apple Silicon. Supported DFlash speculative decoding and image input.
  • August 2026 Meta Superintelligence Lab released Muse Glimmer-30B under the Apache 2.0 license. They announced that it is a model distillation from Muse Spark and is designed primarily for local autonomous agent use.

5. Real-world performance (Community reception)

Insufficient information has been gathered yet.

6. Recommended parameters

The generation parameters recommended in the Best Practices section of the model card are as follows:

Parameter Recommended Value
temperature 1.0
top_p 0.95
top_k 64

Additionally, by writing Reasoning strength: <value> in the system prompt, the inference strength can be selected from four levels: low / medium / high / xhigh. High or xhigh is recommended for complex problem solving, coding, and agent tasks.

Source: Best Practices section of the Hugging Face model card "Muse Glimmer Model Card."

7. Sources