Qwen3-4B-Base: benchmark results
Alibaba's Qwen3-4B-Base, the pretrained (non-instruct) dense 4B checkpoint in the Qwen3 series, trained on 36T tokens across 119 languages with a three-stage pretraining pipeline. Provider: Alibaba. Released 2025-04-29. Access: Open.
Aggregate Rating 1467 ± 17, rank #885 of 1635 rated models, from 216 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| STEMVerse - Chemistry - Chemical Engineering and Technology | 34.04 | Few-shot accuracy (%) | 100 |
| STEMVerse - Chemistry - Theoretical and Computational Chemistry | 33.77 | Few-shot accuracy (%) | 100 |
| STEMVerse - Chemistry - Bloom Level: Apply | 24.93 | Few-shot accuracy (%) | 93.3 |
| Swallow - Pre-trained English - HumanEval | 61.7 | Pass@1 (%) | 89.8 |
| Open Arabic LLM - Arabic MMLU Philosophy (High School) | 69.23 | Accuracy (%) | 88.6 |
| Open Arabic LLM - Aratrust MentalHealth | 94.74 | Accuracy (%) | 87.3 |
| STEMVerse - Chemistry - Analytical Chemistry | 27.43 | Few-shot accuracy (%) | 86.7 |
| STEMVerse - Chemistry - Physical Chemistry | 37.25 | Few-shot accuracy (%) | 86.7 |
| Swallow - Pre-trained English - MATH | 52 | Exact Match (%) | 84.7 |
| Swallow - Pre-trained Japanese - JHumanEval | 53.7 | Pass@1 (%) | 83.7 |
| Open Arabic LLM - Arabic MMLU HT Abstract Algebra | 38 | Accuracy (%) | 81.8 |
| Open Arabic LLM - Aratrust Illegal | 92.45 | Accuracy (%) | 81.7 |
Interactive version: theaggregate.ai/model?slug=qwen3-4b-base · How It Works · Data refreshed daily, snapshot 2026-10-15.