Qwen 3 8B (Thinking): benchmark results
Qwen 3 8B evaluated with thinking enabled. Provider: Alibaba. Released 2025-04-29. Access: Open.
Aggregate Rating 1504 ± 6, rank #1071 of 2170 rated models, from 489 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| GeoBenchLLM - GKMC | 82 | Accuracy (%) | 100 |
| GeoBenchLLM - GeoSQA | 63 | Accuracy (%) | 100 |
| GeoBenchLLM - GridRoute | 81 | Optimal route ratio (%) | 100 |
| GeoBenchLLM - PPNL Multi-Goal | 62 | Optimal route ratio (%) | 100 |
| GeoBenchLLM - PPNL Single-Goal | 92 | Optimal route ratio (%) | 100 |
| LLM Self-Modeling - Confidence-Recall | -0.08 | Skill above dummy predictor (-1 to 1, higher is better; 4096 | 100 |
| LLM Self-Modeling - Edit-Proposal | 0.57 | Skill above dummy predictor (-1 to 1, higher is better; 4096 | 100 |
| LLM Self-Modeling - Feature-Rate | 0.05 | Skill above dummy predictor (-1 to 1, higher is better; 4096 | 100 |
| LLM Self-Modeling - Output-Prediction | 0.49 | Skill above dummy predictor (-1 to 1, higher is better; 4096 | 100 |
| LLM Self-Modeling - Perturbation-Choice | 0.07 | Skill above dummy predictor (-1 to 1, higher is better; 4096 | 100 |
| GT-HarmBench - Stag Hunt | 85 | Utilitarian accuracy (%) on the 317 Stag Hunt scenarios: sha | 92.9 |
| Werewolf Accusation Benchmark - Role Separation | 1.51 | Belief separation (wolves minus villagers, -3..+3 scale) | 87.2 |
Interactive version: theaggregate.ai/model?slug=qwen-3-8b-thinking · How It Works · Data refreshed daily, snapshot 2026-11-05.