Qwen 3.6 Flash: benchmark results
Alibaba's Qwen 3.6 Flash (April 2026), a native vision-language MoE model tuned for agentic coding and code reasoning, part of the Qwen3.6 generation focused on real-world agent workflows. Provider: Alibaba. Released 2026-04-27. Access: API.
Aggregate Rating 1692 ± 9, rank #214 of 1656 rated models, from 313 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| HOLMES | 59.54 | Overall Accuracy (self-reported) | 100 |
| UAV-DualCog - Landmark-Relative Direction | 62.7 | Answer accuracy (%; 1,024 questions asking where a landmark | 97.1 |
| ReaLM - Clean Accuracy | 71.7 | Accuracy (%) on the unperturbed images of the 832 ReaLM phys | 91.7 |
| ReaLM - PhysPatch | 29.1 | Attack success rate (%; share of adversarial cases where the | 91.7 |
| UAV-DualCog - Landmark-Driven Action Decision | 56.7 | Answer accuracy (%; 1,024 questions asking which way the UAV | 91.4 |
| THOR Finding Triage - False Review Load | 29.9 | False positives not suppressed (%, lower is better) | 90.6 |
| Horangi 4 - HAE-RAE Bench (Reading Comprehension) | 91 | Accuracy (%) | 88 |
| Evals for Every Language - Mari (chm) | 61.66 | Average Score (%) | 85.5 |
| Horangi 4 - KorQuAD 1.0 | 86.97 | F1 (x100) | 84.6 |
| AI for Education Visual Maths - Measurement | 94.59 | Accuracy (%) | 84.5 |
| BLXBench | 82.8 | Score (imported) | 83.3 |
| Evals for Every Language - Limburgish (li) | 74.5 | Average Score (%) | 82.8 |
Interactive version: theaggregate.ai/model?slug=qwen-3-6-flash · How It Works · Data refreshed daily, snapshot 2026-10-20.