Qwen 3 4B 2507 Instruct: benchmark results
Dedicated direct-answer checkpoint from the July Qwen3 update. Does not emit thinking blocks; post-training targets instruction following, tool use and multilingual knowledge. Provider: Alibaba. Released 2025-08-06. Access: Open.
Aggregate Rating 1506 ± 8, rank #762 of 1656 rated models, from 692 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| Granite Arabic Card - Question Answering - Average | 67.16 | Exact Match (%) | 100 |
| Granite Arabic Card - Question Answering - Belebele (En) | 91.67 | Exact Match (%) | 100 |
| Granite Arabic Card - Question Answering - Belebele (MSA) | 85.22 | Exact Match (%) | 100 |
| Granite Arabic Card - Question Answering - DialArMMLU (En) | 75.41 | Exact Match (%) | 100 |
| Granite Arabic Card - Question Answering - MMLU-ProX (En) | 60.77 | Exact Match (%) | 100 |
| Granite Arabic Card - Question Answering - MMLU-ProX (MSA) | 46.17 | Exact Match (%) | 100 |
| Industrial-Instruction Panasonic QA | 28.5 | Set-match accuracy (%; share of the 1,000 held-out multi-sel | 100 |
| Kurakurai French LLM - AIME 2025 | 38 | Accuracy (%) | 100 |
| Kurakurai French LLM - MATH-500 | 92.46 | Accuracy (%) | 100 |
| Kurakurai French LLM - MBPP | 76.03 | pass@1 (%) | 100 |
| Kurakurai French LLM - MGSM | 93.08 | Accuracy (%) | 100 |
| Kurakurai French LLM - MultiIF | 71.35 | Accuracy (%) | 100 |
Interactive version: theaggregate.ai/model?slug=qwen-3-4b-2507-instruct · How It Works · Data refreshed daily, snapshot 2026-10-20.