Qwen 3.8 27B: benchmark results
Alibaba's open-weight Qwen3.8-27B (August 2026), a compact dense vision-language model in the Qwen3.8 family, built on the Qwen3.5 architecture with flexible thinking control for coding and agentic tasks. Provider: Alibaba. Released 2026-08-14. Access: Open.
Aggregate Rating 1759 ± 6, rank #105 of 1656 rated models, from 719 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| Context Arena | 93.98 | Average Score (%) | 100 |
| Game2Real-Bench - Real World - Fine-grained Perception | 67 | Task score (%) | 100 |
| HAE-GEO - Fooled Rate (L1) | 26.6 | Fooled Rate (%) | 100 |
| HAE-GEO - Fooled Rate (L3) | 31.7 | Fooled Rate (%) | 100 |
| Jais 3 Model Cards - GAIA v2 lit (Arabic) | 55.8 | Pass@1 (%) | 100 |
| Jais 3 Model Cards - GDPval (75% Threshold) | 78.2 | Accuracy (%) | 100 |
| PureDocBench - Real Degraded - Text Edit Distance | 27.4 | Text normalized edit distance (%): edit distance between the | 100 |
| YKS Exam 2018 - Total | 189 | Net Points (out of 200) | 100 |
| YKS Exam 2019 - TYT | 115 | Net Points (out of 120) | 100 |
| YKS Exam 2019 - TYT Turkish | 36 | Net Points (out of 40) | 100 |
| YKS Exam 2019 - Total | 195 | Net Points (out of 200) | 100 |
| YKS Exam 2021 - TYT Mathematics | 40 | Net Points (out of 40) | 100 |
Interactive version: theaggregate.ai/model?slug=qwen-3-8-27b · How It Works · Data refreshed daily, snapshot 2026-10-20.