Qwen 3.5 4B (Non-reasoning): benchmark results
Qwen 3.5 4B evaluated with reasoning disabled. Provider: Alibaba. Released 2026-03-02. Access: Open.
Aggregate Rating 1495 ± 12, rank #1094 of 2168 rated models, from 144 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| MCJudgeBench - Response-Variant Inconsistency | 3.59 | Procedural inconsistency rate under candidate-response varia | 100 |
| DistortBench - Compression | 56.4 | Joint accuracy (%) on the two compression types (JPEG, JPEG2 | 97.1 |
| PRISM-VLM - Counting | 64.6 | Per-axis mean score (%; exact-integer accuracy on synthesize | 95.1 |
| DistortBench - Blur | 65.5 | Joint accuracy (%) on the three blur types (Gaussian, lens, | 88.2 |
| PRISM-VLM - Multi-Question Robustness | 48.5 | Per-axis mean score (%; share of items whose 3-5 bundled que | 82.9 |
| PRISM-VLM - Quality | 68.2 | Per-axis mean score (%; base accuracy on the recycled questi | 82.9 |
| DistortBench | 56.1 | Joint accuracy (%) on DistortBench's 13,500 four-choice ques | 82.4 |
| DistortBench - Spatial | 58.1 | Joint accuracy (%) on the eight spatial distortions (jitter, | 82.4 |
| AA TAU-2 Bench | 87.72 | Accuracy (%) | 81.9 |
| Tau2-Bench Telecom | 87.72 | Success Rate (%) | 81.7 |
| ObGynLongBench - History-Level EHR - L3 Long-Horizon Reasoning | 59.3 | Accuracy (%) | 81.2 |
| PRISM-VLM | 59.8 | PScore (%; harmonic mean of the seven per-axis mean scores, | 77.5 |
Interactive version: theaggregate.ai/model?slug=qwen-3-5-4b-non-reasoning · How It Works · Data refreshed daily, snapshot 2026-10-20.