Model benchmarks
mtplx-qwen38-27b-optimized-speed local LLM performance
As of October 2026, mtplx-qwen38-27b-optimized-speed runs at up to 45.5 tok/s for local inference (best of 4 community benchmark runs across 3 GPUs).
Model size
27B
Peak speed
45.5 tok/s
Average speed
33.1 tok/s
Avg PP
115.5 tok/s
Min memory
28.1 GB
Max context
128,000 tokens
Avg output / run
48,843 tokens
Avg runtime / run
21m 46s
Avg quality
85.0
Benchmark runs
4
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 4 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 | 83.7 | 85.0 | 86.3 |
| Agent Workflow | 86.2 | 88.6 | 91.1 |
| Code Generation | 71.4 | 76.9 | 83.0 |
| Role Play & Narrative | 83.6 | 89.0 | 94.1 |
| Research & Analysis | 83.0 | 85.5 | 88.3 |
Performance by hardware and tool
Every hardware/tool/quantization combination mtplx-qwen38-27b-optimized-speed 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 M4 Max | vLLM | — | 45.5 tok/s | 45.5 tok/s | n/a | 128,000 tokens | 86.4 | 1 |
| Apple M5 Max | vLLM | — | 40.8 tok/s | 40.8 tok/s | n/a | 128,000 tokens | 85.8 | 1 |
| Apple M4 Pro | MTPLX | — | 23.3 tok/s | 23.1 tok/s | 28.1 GB | 16,384 tokens | 84.0 | 2 |
Benchmark runs
All 4 mtplx-qwen38-27b-optimized-speed runs submitted so far — expand one for its hardware, quality breakdown and per-scenario detail.
Frequently asked questions
- Is mtplx-qwen38-27b-optimized-speed good for coding?
- In our benchmarks, mtplx-qwen38-27b-optimized-speed scores 77.0/100 for coding. It runs at about 33.1 tok/s, so if you want more speed, local-qwen is faster (~60.2 tok/s) and still scores well for coding (84.8/100).
- Is mtplx-qwen38-27b-optimized-speed good for agentic (tool-using) tasks?
- In our benchmarks, mtplx-qwen38-27b-optimized-speed scores 88.6/100 for agentic workflows. It runs at about 33.1 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 mtplx-qwen38-27b-optimized-speed for local inference?
- Across 4 community benchmark runs, mtplx-qwen38-27b-optimized-speed reaches up to 45.5 tok/s and averages 33.1 tok/s, with the fastest results on Apple M4 Max.
- How much memory does mtplx-qwen38-27b-optimized-speed need?
- The leanest observed configuration used about 28.1 GB of memory.
- Which tools have been used to run mtplx-qwen38-27b-optimized-speed?
- Benchmarks were submitted using MTPLX, vLLM. Results are community-contributed and updated as new runs arrive.