InternVL2.5-78B
OpenGVLab · 2024-12-05 · 78.4B parameters
InternVL2.5-78B is OpenGVLab's largest InternVL 2.5 vision-language checkpoint, combining a 6B InternViT vision encoder with Qwen2.5-72B-Instruct. The downloadable checkpoint is governed by the Qwen License; the surrounding InternVL project code has separate terms.
Benchmark scores
| Benchmark | Score |
|---|---|
| AI2D (OpenVLM) | 89.1 |
| Capability | 128.7 |
| CharXiv Reasoning | 42.4 |
| CharXiv-D | 82.3 |
| HallusionBench (OpenVLM) | 57.4 |
| MATH-Vision | 32.2 |
| MathVista (OpenVLM) | 70.6 |
| MM-Vet (OpenVLM) | 16.9 |
| MMMU Validation (OpenVLM) | 70.0 |
| MMStar (OpenVLM) | 69.5 |
| Video-MME | 72.1 |