Model benchmarks
Qwen3.8-27B-oQ8e-fp16-mtp local LLM performance
As of October 2026, Qwen3.8-27B-oQ8e-fp16-mtp runs at up to 35.5 tok/s for local inference (best of 18 community benchmark runs across 3 GPUs).
Model size
27B
Peak speed
35.5 tok/s
Average speed
33.0 tok/s
Avg PP
229.6 tok/s
Min memory
33.3 GB
Max context
262,144 tokens
Avg output / run
25,273 tokens
Avg runtime / run
15m 59s
Avg quality
82.0
Benchmark runs
18
GPUs tested
3
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 18 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 | 69.5 | 82.0 | 88.2 |
| Agent Workflow | 61.6 | 80.2 | 90.8 |
| Code Generation | 53.0 | 73.7 | 84.4 |
| Role Play & Narrative | 78.9 | 89.0 | 94.7 |
| Research & Analysis | 78.4 | 84.9 | 89.3 |
Performance by hardware and tool
Every hardware/tool/quantization combination Qwen3.8-27B-oQ8e-fp16-mtp has been benchmarked on, ranked by peak token generation speed. Last updated October 2026.
| Hardware | Tool | Quant | Best tok/s | Avg tok/s | Memory | Context | Quality | Runs |
|---|---|---|---|---|---|---|---|---|
| Apple M2 Ultra | oMLX | FP16 | 35.5 tok/s | 34.1 tok/s | 59.1 GB | 262,144 tokens | 80.7 | 13 |
| Apple M5 Max | oMLX | FP16 | 35.4 tok/s | 33.7 tok/s | 33.3 GB | 262,144 tokens | 84.6 | 4 |
| Apple M1 Max | oMLX | FP16 | 16.2 tok/s | 16.2 tok/s | 33.8 GB | 65,536 tokens | 88.2 | 1 |
Benchmark runs
All 18 Qwen3.8-27B-oQ8e-fp16-mtp runs submitted so far — expand one for its hardware, quality breakdown and per-scenario detail.
Frequently asked questions
- Is Qwen3.8-27B-oQ8e-fp16-mtp good for coding?
- In our benchmarks, Qwen3.8-27B-oQ8e-fp16-mtp scores 73.7/100 for coding. It runs at about 33.0 tok/s, so if you want more speed, byteshape/Qwen3.8-27B-GGUF:IQ4_XS-ASCII-Condensed is faster (~42.0 tok/s) and still scores well for coding (86.3/100).
- Is Qwen3.8-27B-oQ8e-fp16-mtp good for agentic (tool-using) tasks?
- In our benchmarks, Qwen3.8-27B-oQ8e-fp16-mtp scores 80.2/100 for agentic workflows. It runs at about 33.0 tok/s, so if you want more speed, granite-4.2-8b-Q4_K_L is faster (~71.6 tok/s) and still scores well for agentic workflows (92.9/100).
- How fast is Qwen3.8-27B-oQ8e-fp16-mtp for local inference?
- Across 18 community benchmark runs, Qwen3.8-27B-oQ8e-fp16-mtp reaches up to 35.5 tok/s and averages 33.0 tok/s, with the fastest results on Apple M2 Ultra.
- How much memory does Qwen3.8-27B-oQ8e-fp16-mtp need?
- The leanest observed configuration used about 33.3 GB of memory (quantizations tested: FP16).
- Which tools have been used to run Qwen3.8-27B-oQ8e-fp16-mtp?
- Benchmarks were submitted using oMLX. Results are community-contributed and updated as new runs arrive.