Qwen 3.5 4B (Thinking): benchmark results
Provider: Alibaba. Released 2026-03-02. Access: Open.
Unified ELO 1533 ± 1, rank #1107 of 3078 rated models, from 108 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| Nejumi 4 - BFCL - Live AST | 74.07 | Accuracy (%) | 89.3 |
| Nejumi 4 - BFCL - Relevance Detection | 88.89 | Accuracy (%) | 89 |
| Swallow - Post-trained English - GPQA Diamond | 76.5 | Accuracy (%) | 83.6 |
| Swallow - Post-trained Japanese - GPQA | 68.8 | Accuracy (%) | 77.6 |
| FrameBench - Frame Identification - English | 81 | Accuracy (%; FrameNet candidate frames) | 76.3 |
| Nejumi 4 - Toxicity - Prohibited Acts | 95.83 | Criteria met (%) | 75.4 |
| Swallow - Post-trained English - MMLU-Pro | 78.8 | Accuracy (%) | 71.6 |
| Nejumi 4 - HalluLens | 95 | Hallucination resistance (%) | 71.3 |
| Nejumi 4 - Toxicity - Violation Categories | 46.83 | Criteria met (%) | 71 |
| Nejumi 4 - JHumanEval | 39.03 | Sandbox pass rate (%) | 69.9 |
| Nejumi 4 - Toxicity - Fairness | 97.18 | Criteria met (%) | 69.5 |
| Nejumi 4 - jaster (0-shot) - MAWPS | 98 | Exact match (%) | 68.4 |
Interactive version: theaggregate.ai/model?slug=qwen-3-5-4b-thinking · How It Works · Data refreshed daily, snapshot 2026-09-19.