Model benchmarks
qwen3.8:27b-mlx local LLM performance
As of November 2026, qwen3.8:27b-mlx runs at up to 40.3 tok/s for local inference (best of 10 community benchmark runs across 5 GPUs).
Model size
27.8B
Peak speed
40.3 tok/s
Average speed
24.2 tok/s
Avg PP
1881.3 tok/s
Min memory
20.4 GB
Max context
128,000 tokens
Avg output / run
34,719 tokens
Avg runtime / run
18m 48s
Avg quality
81.3
Benchmark runs
10
GPUs tested
5
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 10 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 | 72.8 | 81.3 | 86.3 |
| Agent Workflow | 72.7 | 81.2 | 88.0 |
| Code Generation | 33.9 | 68.4 | 79.5 |
| Role Play & Narrative | 85.0 | 91.5 | 95.3 |
| Research & Analysis | 79.5 | 84.3 | 87.5 |
Performance by hardware and tool
Every hardware/tool/quantization combination qwen3.8:27b-mlx has been benchmarked on, ranked by peak token generation speed. Last updated November 2026.
| Hardware | Tool | Quant | Best tok/s | Avg tok/s | Memory | Context | Quality | Runs |
|---|---|---|---|---|---|---|---|---|
| Apple M5 Max | Ollama | nvfp4 | 40.3 tok/s | 35.1 tok/s | 20.8 GB | 128,000 tokens | 83.2 | 2 |
| Apple M5 Pro | Ollama | nvfp4 | 36.7 tok/s | 25.1 tok/s | 21.3 GB | 32,768 tokens | 80.4 | 2 |
| Apple M4 Pro | Ollama | nvfp4 | 28.6 tok/s | 22.5 tok/s | 20.4 GB | 32,768 tokens | 81.0 | 2 |
| Apple M5 Max | LM Studio | — | 22.2 tok/s | 22.2 tok/s | n/a | 32,768 tokens | 67.5 | 1 |
| Apple M1 Max | Ollama | nvfp4 | 22.1 tok/s | 21.1 tok/s | 22.7 GB | 65,536 tokens | 85.8 | 2 |
| Apple M2 Pro | Ollama | nvfp4 | 12.3 tok/s | 12.3 tok/s | 20.5 GB | 65,536 tokens | 85.0 | 1 |
Benchmark runs
All 10 qwen3.8:27b-mlx runs submitted so far — expand one for its hardware, quality breakdown and per-scenario detail.
Frequently asked questions
- Is qwen3.8:27b-mlx good for coding?
- In our benchmarks, qwen3.8:27b-mlx scores 68.4/100 for coding. It runs at about 24.2 tok/s, so if you want more speed, Qwen3.8-27B-Turbo-Brilliance-50X-RI-GSQ-RCO-IQ3_XXS-mtp is faster (~37.5 tok/s) and still scores well for coding (83.8/100).
- Is qwen3.8:27b-mlx good for agentic (tool-using) tasks?
- In our benchmarks, qwen3.8:27b-mlx scores 81.2/100 for agentic workflows. It runs at about 24.2 tok/s, so if you want more speed, muse-glimmer:latest is faster (~33.3 tok/s) and still scores well for agentic workflows (91.6/100).
- How fast is qwen3.8:27b-mlx for local inference?
- Across 10 community benchmark runs, qwen3.8:27b-mlx reaches up to 40.3 tok/s and averages 24.2 tok/s, with the fastest results on Apple M5 Max.
- How much memory does qwen3.8:27b-mlx need?
- The leanest observed configuration used about 20.4 GB of memory (quantizations tested: nvfp4).
- Which tools have been used to run qwen3.8:27b-mlx?
- Benchmarks were submitted using LM Studio, Ollama. Results are community-contributed and updated as new runs arrive.