Qwen 3.5 9B: benchmark results
Alibaba's open 9B dense model of the Qwen3.5 small series: hybrid linear attention, native multimodality, 262K context. Provider: Alibaba. Released 2026-03-02. Access: Open.
Aggregate Rating 1622 ± 3, rank #379 of 1656 rated models, from 2332 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| AgentViSS - Interaction Regulation (Direct Vision) | 40.9 | Interaction-regulation task score, the images given directly | 100 |
| CBX-Bench (Judge Alignment) | 63.3 | Ranking-recovery alignment (%; share of human pairwise CBM-e | 100 |
| COMEX - Explanation METEOR | 0.27 | METEOR (0-1, native; similarity of the generated composition | 100 |
| CanLegalRAGBench (Oracle Retrieval) - Answer Accuracy | 76.46 | Claim accuracy (%; share of the answer's atomic claims entai | 100 |
| Chinese Social Pragmatic Inference - Comment Only | 33.58 | Leave-writer-out Accuracy (%) | 100 |
| Chinese Social Pragmatic Inference - Full Context | 81.42 | Leave-writer-out Accuracy (%) | 100 |
| ClashBench - Successful Preemption Rate (OpenCode) | 32.1 | SPR (%) | 100 |
| DeepTumorVQA - Recognition | 58.4 | Accuracy (%) on the Recognition-stage questions (binary: les | 100 |
| DeepTumorVQA Agent (Vision Tools) | 40.3 | Answer accuracy (%) on the balanced 10,000-question multiple | 100 |
| Dr-CiK - Codex Evidence | 0.3 | Scaled CRPS (mean over 225 tasks, per-task values winsorized | 100 |
| EMO Reference - AIME 2025 and 2026 (greedy zero-shot) | 25 | Accuracy (%) | 100 |
| EuroEval Catalan NLU - Guia CAT | 77.18 | Matthews correlation coefficient (%) | 100 |
Interactive version: theaggregate.ai/model?slug=qwen-3-5-9b · How It Works · Data refreshed daily, snapshot 2026-10-20.