InternVL3-78B
OpenGVLab · 2025-04-11 · 78.4B parameters
InternVL3-78B is OpenGVLab's 78B-parameter vision-language checkpoint built with a Qwen2.5-72B language component. The downloadable model weights are governed by the Qwen License; the surrounding InternVL project code is separately released under MIT.
Benchmark scores
| Benchmark | Score |
|---|---|
| AI2D (OpenVLM) | 89.8 |
| Capability | 133.5 |
| CharXiv Reasoning | 46.0 |
| CharXiv-D | 85.1 |
| HallusionBench (OpenVLM) | 59.1 |
| MATH-Vision | 43.1 |
| MathVista (OpenVLM) | 79.0 |
| MM-Vet (OpenVLM) | 80.7 |
| MMMU Validation (OpenVLM) | 72.2 |
| MMStar (OpenVLM) | 73.4 |
| OmniDocBench v1.5 | 80.3 |
| SnakeBench Average Apples | 1.2 |
| SnakeBench Best Apples | 7.0 |
| SnakeBench Rating | 19.4 |
| SnakeBench Total Apples | 51.0 |
| SnakeBench Win Rate | 43.2 |