Qwen 3.5 9B (Non-reasoning): benchmark results
Qwen 3.5 9B evaluated with reasoning disabled. Provider: Alibaba. Released 2026-03-02. Access: Open.
Aggregate Rating 1533 ± 8, rank #944 of 2168 rated models, from 283 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| Know2Guess | 57.04 | Reliability (0-1 scaled to %): the share of test items handl | 100 |
| Know2Guess - Answer Accuracy | 48.33 | Answer accuracy (0-1 scaled to %) on the answer-expected zon | 100 |
| PRISM-VLM - Counting | 66.4 | Per-axis mean score (%; exact-integer accuracy on synthesize | 100 |
| SVFSearch - Plan-Act-Replan Agent | 79.1 | Accuracy (%) on the 5,000-item SVFSearch test split (four-ch | 100 |
| PRISM-VLM - Quality | 69.9 | Per-axis mean score (%; base accuracy on the recycled questi | 95.1 |
| RobustBench-TC - Misleading Description (Abbreviated Name) | 13.2 | Tool-call accuracy (%) under the MisDesc-Abbr perturbation ( | 94.1 |
| RobustBench-TC - Timeout | 71.4 | Tool-call accuracy (%) under the Timeout perturbation (trans | 94.1 |
| PRISM-VLM - Multi-Question Robustness | 50.5 | Per-axis mean score (%; share of items whose 3-5 bundled que | 92.7 |
| PointQ-Bench - Defect Diagnosis | 39.95 | Sample-level F1 (%) of the multi-label defect diagnosis (whi | 92.3 |
| FindIt - Multi-Label Object Detection | 60.9 | Average F1@0.5 (%) over multi-label object detection (every | 88.9 |
| DistortBench - Compression | 55.1 | Joint accuracy (%) on the two compression types (JPEG, JPEG2 | 88.2 |
| RobustBench-TC - Misleading Description | 41.2 | Tool-call accuracy (%) under the MisDesc perturbation (rewar | 88.2 |
Interactive version: theaggregate.ai/model?slug=qwen-3-5-9b-non-reasoning · How It Works · Data refreshed daily, snapshot 2026-10-20.