Model benchmarks
qwen3.8-27b@iq2_xxs local LLM performance
As of September 2026, qwen3.8-27b@iq2_xxs runs at up to 17.5 tok/s for local inference (best of 2 community benchmark runs across 2 GPUs).
Model size
27B
Peak speed
17.5 tok/s
Average speed
10.3 tok/s
Avg prefill
385.1 tok/s
Min memory
8.5 GB
Max context
115,000 tokens
Avg output / run
31,543 tokens
Avg runtime / run
32m 18s
Avg quality
38.0
Benchmark runs
2
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 2 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 | 3.8 | 38.0 | 72.1 |
| Agent Workflow | 4.4 | 44.2 | 84.0 |
| Code Generation | 3.1 | 30.6 | 58.1 |
| Role Play & Narrative | 3.5 | 35.4 | 67.2 |
| Research & Analysis | 4.2 | 41.7 | 79.1 |
Performance by hardware and tool
Every hardware/tool/quantization combination qwen3.8-27b@iq2_xxs 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 9070/9070 XT/9070 GRE | LM Studio | — | 17.5 tok/s | 17.5 tok/s | 9.3 GB | 115,000 tokens | 75.9 | 1 |
| NVIDIA GeForce RTX 4060 Laptop GPU | Ollama | Q4_K_S | 3.1 tok/s | 3.1 tok/s | 8.5 GB | 512 tokens | 0.0 | 1 |
Frequently asked questions
- Is smtek/Qwen3.8-27B:IQ2_XXS good for coding?
- In our benchmarks, smtek/Qwen3.8-27B:IQ2_XXS scores 30.6/100 for coding. It runs at about 10.3 tok/s, so if you want more speed, K2-Horizon-MoVA-36B-A4B-MLX-4bit is faster (~48.8 tok/s) and still scores well for coding (85.3/100).
- Is smtek/Qwen3.8-27B:IQ2_XXS good for agentic (tool-using) tasks?
- In our benchmarks, smtek/Qwen3.8-27B:IQ2_XXS scores 44.2/100 for agentic workflows. It runs at about 10.3 tok/s, so if you want more speed, muse-glimmer:latest is faster (~34.5 tok/s) and still scores well for agentic workflows (93.0/100).
- How fast is qwen3.8-27b@iq2_xxs for local inference?
- Across 2 community benchmark runs, qwen3.8-27b@iq2_xxs reaches up to 17.5 tok/s and averages 10.3 tok/s, with the fastest results on AMD Radeon RX 9070/9070 XT/9070 GRE.
- How much memory does qwen3.8-27b@iq2_xxs need?
- The leanest observed configuration used about 8.5 GB of memory (quantizations tested: Q4_K_S).
- Which tools have been used to run qwen3.8-27b@iq2_xxs?
- Benchmarks were submitted using LM Studio, Ollama. Results are community-contributed and updated as new runs arrive.