QwQ-32B: benchmark results
Alibaba's RL-trained 32B reasoning model, matching the far larger DeepSeek-R1 on math and coding under Apache 2.0 (March 2025). Provider: Alibaba. Released 2025-03-05. Access: Open.
Aggregate Rating 1606 ± 6, rank #421 of 1656 rated models, from 688 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| AgentJudgeBench (Hard, With Ground Truth) - Llama-3.1-8B-Instruct Outputs | 84.1 | Judge alignment (%; agreement of the judge's per-metric verd | 100 |
| AgentJudgeBench (Hard, With Ground Truth) - Llama-3.3-70B-Instruct Outputs | 87.5 | Judge alignment (%; agreement of the judge's per-metric verd | 100 |
| AgentJudgeBench (Hard, With Ground Truth) - Qwen3-32B Outputs | 84.7 | Judge alignment (%; agreement of the judge's per-metric verd | 100 |
| AppellateGen | 2.46 | Mean LLM-Judge Score (0-5; verdict, fact, legal application, | 100 |
| AstroMind - Threat Assessment | 66.67 | Accuracy (%; 92 open-ended threat-level assessments checked | 100 |
| IBA-Bench - Daily Consumption | 80.6 | Checkpoint satisfaction (%; share of hidden preference check | 100 |
| PERSPECTRA - Opinion Counting | 36.2 | Exact-match accuracy (%) of the predicted number of distinct | 100 |
| PilotBench - Regression Score | 85.53 | Regression score (0-100): 0.6 x piecewise-linear MAE score p | 100 |
| RoundTripCodeEval - Arithmetic Coding | 15.71 | Mean of exact match, edit similarity and pass@5 (%) over the | 100 |
| RoundTripCodeEval - Huffman Coding | 5.5 | Mean of exact match, edit similarity and pass@5 (%) over the | 100 |
| RoundTripCodeEval - LZW | 24.14 | Mean of exact match, edit similarity and pass@5 (%) over the | 100 |
| RoundTripCodeEval - Run-Length Encoding | 57.23 | Mean of exact match, edit similarity and pass@5 (%) over the | 100 |
Interactive version: theaggregate.ai/model?slug=qwq-32b · How It Works · Data refreshed daily, snapshot 2026-10-20.