Qwen 3 235B A22B: benchmark results
Can switch between direct answers and explicit reasoning within the same conversation, without loading another checkpoint. Training covers multilingual dialogue, code and tool use. Provider: Alibaba. Released 2025-04-29. Access: Open.
Aggregate Rating 1649 ± 4, rank #313 of 1656 rated models, from 1169 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| AgentSocialBench - Cross-User Leakage | 0.14 | Privacy leakage rate (0-1, share of designated private items | 100 |
| AgentSocialBench - Hub-and-Spoke Leakage | 0.06 | Privacy leakage rate (0-1, share of designated private items | 100 |
| CASTLE Student Safety (English) - Learning Dependence and Cognition | 1.91 | Average Safety Score (1-5) on CASTLE's English scenarios, th | 100 |
| DiagFlowBench - Termination Recognition | 98.5 | Termination recognition (%): share of terminal-node turns at | 100 |
| EComStage - Query Match | 98.81 | Task Score (%) | 100 |
| EmoHarbor | 4.13 | Mean of 10 Rubric Dimensions (1-5 Likert) | 100 |
| ForeSci - Prediction Factuality - Bottleneck-Opportunity Discovery | 84 | Prediction factuality (claim-level F1 x 100; answer claims s | 100 |
| ForeSci - Prediction Factuality - Strategic Research Planning | 46.5 | Prediction factuality (claim-level F1 x 100; answer claims s | 100 |
| GABench - Graph ML - Node Classification (Text-Attributed Graphs) | 22.23 | Task success rate (%; share of tasks whose final answer a De | 100 |
| GABench - Open-Ended QA - Prediction (Numerical-Attribute Graphs) | 4.36 | Task success rate (%; share of tasks whose final answer a De | 100 |
| LogicScore - 2WikiMultiHopQA Completeness | 86 | Completeness (%; answers whose extracted propositions link t | 100 |
| LogicScore - 2WikiMultiHopQA Determinateness | 96.74 | Determinateness (%; answers whose own short answer o4-mini r | 100 |
Interactive version: theaggregate.ai/model?slug=qwen-3-235b-a22b · How It Works · Data refreshed daily, snapshot 2026-10-20.