Qwen3-4B places in the top three on 2 benchmarks.
Qwen Apr 2025 Open weights
Its best score on each benchmark, and where that ranks among every model with a score there. We rank each model once, at its best setting.
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| MathIF | 44% | #2 of 12 | Official leaderboard | |
| MadinahQA | 54% | #8 of 39 | Official leaderboard | |
| Vectara Hallucination Leaderboard (2025 dataset) | 5.70% | #22 of 104 | Official leaderboard | |
| Vectara Hallucination Leaderboard (HHEM) | 2.70% | #36 of 148 | Official leaderboard | |
| Open Arabic LLM Leaderboard v2 | 62.9% | #10 of 39 | Official leaderboard | |
| ArbMMLU-HT | 53.7% | #11 of 39 | Official leaderboard | |
| ArabicMMLU | 59.8% | #12 of 39 | Official leaderboard | |
| AlGhafa | 68.5% | #13 of 39 | Official leaderboard | |
| ALRAGE | 76.3% | #15 of 39 | Official leaderboard | |
| AlignBench | 8.30 | #7 of 15 | Lab model card | |
| Arena-Hard-Auto | 76.6% | #20 of 39 | Lab model card | |
| Global PIQA | 70.5% | #23 of 44 | Official leaderboard | |
| Stick to Your Role! | 0.56 | #9 of 15 | Official leaderboard | |
| Multi-IF | 66.3% | #19 of 28 | Lab model card | |
| Alyah | 26.3% | #7 of 10 | Official leaderboard | |
| Slop Score | 62.2 | #12 of 17 | Official leaderboard | |
| RULER | 85.2% | #30 of 42 | Lab model card | |
| IFEval | 81.9% | #99 of 135 | Lab model card | |
| Encyclo-K | 23.6% | #17 of 23 | Official leaderboard | |
| MLogiQA | 65.9% | #6 of 8 | Lab model card | |
| C-Eval | 77.5% | #37 of 46 | Lab model card | |
| INCLUDE | 61.8% | #25 of 31 | Lab model card | |
| MMMLU | 69.8% | #58 of 71 | Lab model card | |
| MMLU-Redux | 83.7% | #37 of 45 | Lab model card | |
| MT-AIME2024 | 60.7% | #6 of 7 | Lab model card | |
| Arena-Hard v2.0 | 15% | #20 of 23 | Official leaderboard | |
| KINA | 21.5% | #33 of 37 | Official leaderboard | |
| Creative Writing v3 | 61.1 | #11 of 12 | Lab model card | |
| WritingBench | 73.5 | #13 of 14 | Lab model card | |
| Longform Creative Writing (EQ-Bench) | 17.0 | #119 of 119 | Official leaderboard | |
| HABIT-Bench | 29.8% | #6 of 6 | Official leaderboard |
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| MATH-500 | 97% | #20 of 70 | Lab model card | |
| A^3-Bench | 58.9% | #2 of 7 | Official leaderboard | |
| IneqMath | 5.50% | #16 of 34 | Official leaderboard | |
| AIME 2024 | 73.8% | #54 of 88 | Lab model card | |
| ZebraLogic | 81% | #15 of 23 | Lab model card | |
| GPQA Diamond | 52.3% | #151 of 229 | Epoch AI run | |
| AutoLogi | 87.9% | #6 of 8 | Lab model card | |
| AIME 2025 | 65.6% | #118 of 143 | Lab model card | |
| PolyMath | 40% | #20 of 24 | Lab model card |
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| LiveCodeBench v5 | 54.2% | #35 of 47 | Lab model card | |
| Codeforces rating | 1,671 | #38 of 47 | Lab model card | |
| StructEval | 67% | #23 of 28 | Official leaderboard | |
| AutoCodeBench | 24.3% | #30 of 34 | Official leaderboard | |
| ArtifactsBench (Gemini 2.5 Pro 0506 judge) | 32.8% | #17 of 18 | Official leaderboard |
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| C3-Bench | 7.81% | #17 of 30 | Official leaderboard | |
| LisanBench | 263.3 | #83 of 133 | Official leaderboard | |
| Berkeley Function Calling Leaderboard (BFCL) | 65.9% | #26 of 41 | Lab model card | |
| SnakeBench | 16 | #164 of 188 | Official leaderboard | |
| LLM Chess | -637 | #115 of 123 | Official leaderboard | |
| EnterpriseOps-Gym | 15.3% | #12 of 12 | Official leaderboard |
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| BizFinBench | 59.9% | #9 of 13 | Official leaderboard | |
| BRIDGE | 42.8 | #28 of 39 | Official leaderboard |
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| AraTrust | 84.9% | #9 of 39 | Official leaderboard | |
| AthenaBench | 37.5% | #6 of 6 | Official leaderboard |
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| LiveBench | 63.6% | #15 of 27 | Lab model card |
The Self-Reported Index fits one score to the 18 capability results Qwen reported for this model, on the scale of the Epoch Capabilities Index. Labs choose which benchmarks to show and how to run them, so read it as the lab's claim, not an independent measurement. The 90% interval runs from 128.0 to 130.9. How we compute it.
Where the benchmark's own leaderboard, or Epoch AI's run, also has this model. Gaps mostly come from different harnesses, effort levels and task subsets. Every pair.
| Benchmark | Lab says | Board says | Gap |
|---|---|---|---|
| GPQA Diamond | 55.9% GPQA-Diamond, thinking mode (Table 17). Non-thinking 41.7 (Table 18) | 52.3% Epoch AI run | +3.6 pts |
| Berkeley Function Calling Leaderboard (BFCL) | 65.9% BFCL v3, thinking mode (Table 17). Non-thinking 57.6 (Table 18) | 63.4% Leaderboard, native function calling | +2.5 pts |
| PolyMath | 40% PolyMath, thinking mode (Table 17). Non-thinking 16.6 (Table 18) | 40% Leaderboard, thinking | Same |
| RULER | 85.2% RULER Avg. (4K to 128K, YaRN x4), non-thinking mode, 128K 66.0. Thinking mode Avg. 83.5 with 8192-token thinking budget (Table 23) | 85.2% Leaderboard | Same |
Every result in the longer documents the lab published with it. 25 of them aren't in the launch post.