Is qwen3.8-27b good for coding?

In our benchmarks, qwen3.8-27b scores 51.4/100 for coding. It runs at about 50.3 tok/s, so if you want more speed, Tiel-Coder-35B-A3B-MLX-oQ4e is faster (~107.9 tok/s) and still scores well for coding (81.5/100).

Is qwen3.8-27b good for agentic (tool-using) tasks?

In our benchmarks, qwen3.8-27b scores 84.7/100 for agentic workflows. It runs at about 50.3 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 qwen3.8-27b for local inference?

Across 26 community benchmark runs, qwen3.8-27b reaches up to 156.8 tok/s and averages 50.3 tok/s, with the fastest results on NVIDIA GeForce RTX 5090.

How much memory does qwen3.8-27b need?

The leanest observed configuration used about 16.3 GB of memory (quantizations tested: Q4_K_M).

Which tools have been used to run qwen3.8-27b?

Benchmarks were submitted using LM Studio, Ollama, llama.cpp, ninfer, vLLM. Results are community-contributed and updated as new runs arrive.

Model benchmarks

qwen3.8-27b local LLM performance

As of September 2026, qwen3.8-27b runs at up to 156.8 tok/s for local inference (best of 26 community benchmark runs across 12 GPUs).

LM StudioOllamallama.cppninfervLLMQ4_K_M

ShareReddit X

Model size

27B

Peak speed

156.8 tok/s

Average speed

50.3 tok/s

Avg PP

1365.2 tok/s

Min memory

16.3 GB

Max context

123,904 tokens

Avg output / run

49,550 tokens

Avg runtime / run

24m 14s

Avg quality

73.1

Benchmark runs

GPUs tested

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 26 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	44.7	73.1	88.0
Agent Workflow	74.3	84.7	91.1
Code Generation	1.6	51.4	82.9
Role Play & Narrative	79.8	89.0	94.5
Research & Analysis	0.3	67.4	90.3

Performance by hardware and tool

Every hardware/tool/quantization combination qwen3.8-27b 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
NVIDIA GeForce RTX 5090	ninfer	—	156.8 tok/s	150.0 tok/s	n/a	65,536 tokens	68.0	2
NVIDIA GeForce RTX 5090	vLLM	—	155.2 tok/s	155.2 tok/s	n/a	65,536 tokens	89.1	1
NVIDIA L4	Ollama	Q4_K_M	109.3 tok/s	109.3 tok/s	16.7 GB	65,536 tokens	86.7	1
AMD Radeon RX 7900 XT/7900 XTX/7900M	llama.cpp	—	52.4 tok/s	51.8 tok/s	n/a	65,536 tokens	58.9	4
AMD Radeon RX 7900 XT/7900 XTX/7900M	LM Studio	—	51.9 tok/s	51.9 tok/s	n/a	65,536 tokens	86.3	1
AMD Radeon RX 7900 XTX	LM Studio	—	51.8 tok/s	51.8 tok/s	n/a	65,536 tokens	45.1	1
NVIDIA GeForce RTX 5070 Ti	LM Studio	—	46.9 tok/s	42.8 tok/s	n/a	121,072 tokens	87.9	2
AMD Radeon RX 7900 XTX	Ollama	Q4_K_M	39.4 tok/s	38.4 tok/s	16.3 GB	65,536 tokens	82.2	4
NVIDIA GeForce RTX 4090	LM Studio	—	39.0 tok/s	39.0 tok/s	n/a	8,192 tokens	66.7	1
AMD Radeon RX 7900 XT/7900 XTX/7900 GRE/7900M	Ollama	Q4_K_M	24.6 tok/s	24.6 tok/s	20.5 GB	65,536 tokens	85.3	1
Apple M5 Max	LM Studio	—	24.3 tok/s	20.8 tok/s	n/a	123,904 tokens	73.0	3
Apple M5 Pro	LM Studio	—	17.1 tok/s	15.2 tok/s	n/a	65,536 tokens	66.7	2
AMD Radeon RX 9070 XT	Ollama	Q4_K_M	15.0 tok/s	15.0 tok/s	16.9 GB	8,192 tokens	85.0	1
Apple M4 Pro	LM Studio	—	14.0 tok/s	14.0 tok/s	n/a	32,768 tokens	85.7	1
Apple M5	LM Studio	—	7.6 tok/s	7.6 tok/s	n/a	8,192 tokens	43.1	1

Benchmark runs

All 26 qwen3.8-27b runs submitted so far — expand one for its hardware, quality breakdown and per-scenario detail.

Open run page →

Frequently asked questions

Is qwen3.8-27b good for coding?: In our benchmarks, qwen3.8-27b scores 51.4/100 for coding. It runs at about 50.3 tok/s, so if you want more speed, Tiel-Coder-35B-A3B-MLX-oQ4e is faster (~107.9 tok/s) and still scores well for coding (81.5/100).
Is qwen3.8-27b good for agentic (tool-using) tasks?: In our benchmarks, qwen3.8-27b scores 84.7/100 for agentic workflows. It runs at about 50.3 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 qwen3.8-27b for local inference?: Across 26 community benchmark runs, qwen3.8-27b reaches up to 156.8 tok/s and averages 50.3 tok/s, with the fastest results on NVIDIA GeForce RTX 5090.
How much memory does qwen3.8-27b need?: The leanest observed configuration used about 16.3 GB of memory (quantizations tested: Q4_K_M).
Which tools have been used to run qwen3.8-27b?: Benchmarks were submitted using LM Studio, Ollama, llama.cpp, ninfer, vLLM. Results are community-contributed and updated as new runs arrive.