InternVL3-38B
OpenGVLab · 2025-04-11 · 38.4B parameters
InternVL3-38B is OpenGVLab's 38,388,164,992-parameter vision-language checkpoint, combining an InternViT-6B-448px-V2.5 encoder with a Qwen2.5-32B language backbone through a two-layer MLP projector. Released after native multimodal pre-training, supervised fine-tuning, and mixed preference optimization, it accepts text, images, and sampled video and generates text; its repository metadata declares Apache-2.0, while the surrounding InternVL project code is MIT.
Benchmark scores
| Benchmark | Score |
|---|---|
| AI2D (OpenVLM) | 88.7 |
| Capability | 132.5 |
| CharXiv Reasoning | 46.4 |
| CharXiv-D | 87.2 |
| HallusionBench (OpenVLM) | 58.4 |
| MATH-Vision | 34.5 |
| MathVista (OpenVLM) | 76.3 |
| MM-Vet (OpenVLM) | 81.1 |
| MMMU Validation (OpenVLM) | 69.7 |
| MMStar (OpenVLM) | 72.6 |