Model benchmarks
muse-glimmer:latest local LLM performance
As of September 2026, muse-glimmer:latest runs at up to 35.1 tok/s for local inference (best of 5 community benchmark runs across 2 GPUs).
Model size
27.9B
Peak speed
35.1 tok/s
Average speed
33.3 tok/s
Avg PP
1167.5 tok/s
Min memory
15.5 GB
Max context
65,536 tokens
Avg output / run
16,270 tokens
Avg runtime / run
7m 50s
Avg quality
83.1
Benchmark runs
5
GPUs tested
2
Quality by task
Average LLM-judged quality (0–100) with the run-to-run spread shown as a P5–P95 band, overall and for each benchmark task, across all 5 runs. The low and high columns show how much the judge’s score varies between runs, and need at least two runs to display.
| Task | P5 (low) | Avg | P95 (high) |
|---|---|---|---|
| Overall | 78.7 | 83.1 | 86.8 |
| Agent Workflow | 87.4 | 91.6 | 93.4 |
| Code Generation | 69.9 | 74.0 | 81.7 |
| Role Play & Narrative | 64.8 | 82.4 | 92.4 |
| Research & Analysis | 81.6 | 84.3 | 86.6 |
Performance by hardware and tool
Every hardware/tool/quantization combination muse-glimmer:latest has been benchmarked on, ranked by peak token generation speed. Last updated September 2026.
| Hardware | Tool | Quant | Best tok/s | Avg tok/s | Memory | Context | Quality | Runs |
|---|---|---|---|---|---|---|---|---|
| AMD Radeon RX 7900 XTX | Ollama | Q4_K_M | 35.1 tok/s | 34.5 tok/s | 15.5 GB | 65,536 tokens | 82.0 | 4 |
| NVIDIA GeForce RTX 4090 | LM Studio | — | 28.3 tok/s | 28.3 tok/s | n/a | 8,192 tokens | 87.5 | 1 |
Benchmark runs
All 5 muse-glimmer:latest runs submitted so far — expand one for its hardware, quality breakdown and per-scenario detail.
Frequently asked questions
- Is muse-glimmer:latest good for coding?
- In our benchmarks, muse-glimmer:latest scores 74.0/100 for coding. It runs at about 33.3 tok/s, so if you want more speed, Qwen3.8-27B-oQ4e-fp16-mtp is faster (~38.6 tok/s) and still scores well for coding (82.6/100).
- Is muse-glimmer:latest good for agentic (tool-using) tasks?
- In our benchmarks, muse-glimmer:latest scores 91.6/100 for agentic workflows, currently the best-scoring model for agentic workflows that runs on consumer hardware (≤24 GB VRAM). It runs at about 33.3 tok/s, so if you want more speed, Nex-N2.5-mini-APEX-Mini is faster (~88.2 tok/s) and still scores well for agentic workflows (91.3/100).
- How fast is muse-glimmer:latest for local inference?
- Across 5 community benchmark runs, muse-glimmer:latest reaches up to 35.1 tok/s and averages 33.3 tok/s, with the fastest results on AMD Radeon RX 7900 XTX.
- How much memory does muse-glimmer:latest need?
- The leanest observed configuration used about 15.5 GB of memory (quantizations tested: Q4_K_M).
- Which tools have been used to run muse-glimmer:latest?
- Benchmarks were submitted using LM Studio, Ollama. Results are community-contributed and updated as new runs arrive.