Qwen3 VL 4B Thinking
Qwen3-VL is a large multimodal model that unifies vision, language, and reasoning to achieve human-level perception and cognition across text, images, and video. Built on a 235B-parameter architecture, it integrates early joint training of visual and textual modalities for strong language grounding. The model supports up to a 1 million-token context window and excels at visual understanding, spatial reasoning, long video comprehension, and tool-based interaction. It can generate code from images, perform precise 2D/3D object grounding, and operate digital interfaces like a visual agent. The “Instruct” version rivals Gemini 2.5 Pro in perception benchmarks, while the “Thinking” version leads in multimodal reasoning and STEM tasks. With multilingual OCR, creative writing, and fine-grained scene interpretation, Qwen3-VL establishes a new open-source frontier for integrated vision-language intelligence.
Benchmark results
| Benchmark | Score | Tags | Source |
|---|---|---|---|
| AI2D | 84.9% | self-reported llm-stats | link → |
| AIME 2025 | 74.5% | self-reported llm-stats | link → |
| Arena-Hard v2 | 36.8% | self-reported llm-stats | link → |
| BFCL-v3 | 67.3% | self-reported llm-stats | link → |
| BLINK | 63.4% | self-reported llm-stats | link → |
| CC-OCR | 73.8% | self-reported llm-stats | link → |
| CharadesSTA | 59.0% | self-reported llm-stats | link → |
| CharXiv-D | 83.9% | self-reported llm-stats | link → |
| CharXiv-R | 50.3% | self-reported llm-stats | link → |
| Creative Writing v3 | 0.761 | self-reported llm-stats | link → |
| DocVQAtest | 94.2% | self-reported llm-stats | link → |
| ERQA | 47.3% | self-reported llm-stats | link → |
| GPQA | 64.1% | self-reported llm-stats | link → |
| Hallusion Bench | 64.1% | self-reported llm-stats | link → |
| HMMT25 | 53.1% | self-reported llm-stats | link → |
| IFEval | 82.6% | self-reported llm-stats | link → |
| Include | 64.6% | self-reported llm-stats | link → |
| InfoVQAtest | 83.0% | self-reported llm-stats | link → |
| LiveBench 20241125 | 68.4% | self-reported llm-stats | link → |
| LiveCodeBench v6 | 51.3% | self-reported llm-stats | link → |
| LVBench | 53.5% | self-reported llm-stats | link → |
| MathVision | 60.0% | self-reported llm-stats | link → |
| MathVista-Mini | 79.5% | self-reported llm-stats | link → |
| MLVU-M | 75.7% | self-reported llm-stats | link → |
| MM-MT-Bench | 7.7 | self-reported llm-stats | link → |
| MMBench-V1.1 | 86.7% | self-reported llm-stats | link → |
| MMLU | 81.5% | self-reported llm-stats | link → |
| MMLU-Pro | 73.6% | self-reported llm-stats | link → |
| MMLU-ProX | 65.0% | self-reported llm-stats | link → |
| MMLU-Redux | 86.0% | self-reported llm-stats | link → |
| MMMU (val) | 70.8% | self-reported llm-stats | link → |
| MMMU-Pro | 57.0% | self-reported llm-stats | link → |
| MMStar | 73.2% | self-reported llm-stats | link → |
| MuirBench | 75.0% | self-reported llm-stats | link → |
| Multi-IF | 73.6% | self-reported llm-stats | link → |
| MVBench | 69.3% | self-reported llm-stats | link → |
| OCRBench | 80.8% | self-reported llm-stats | link → |
| OCRBench-V2 (en) | 61.8% | self-reported llm-stats | link → |
| OCRBench-V2 (zh) | 55.8% | self-reported llm-stats | link → |
| ODinW | 39.4% | self-reported llm-stats | link → |
| OSWorld | 31.4% | self-reported llm-stats | link → |
| PolyMATH | 44.6% | self-reported llm-stats | link → |
| RealWorldQA | 73.2% | self-reported llm-stats | link → |
| ScreenSpot | 92.9% | self-reported llm-stats | link → |
| ScreenSpot Pro | 49.2% | self-reported llm-stats | link → |
| SuperGPQA | 46.8% | self-reported llm-stats | link → |
| VideoMMMU | 69.4% | self-reported llm-stats | link → |
| WritingBench | 84.0% | self-reported llm-stats | link → |