InternVL2-40B
OpenGVLab · 2024-07-08 · 40.1B parameters
InternVL2-40B is OpenGVLab's instruction-tuned vision-language checkpoint released on July 8, 2024. It combines a 45-layer InternViT-6B-448px-V1-5 vision encoder and MLP projector with the 60-layer Nous-Hermes-2-Yi-34B language backbone, accepting text, images, and sampled video frames within an 8,192-token context and generating text. The official repository exposes exactly 40,068,684,160 BF16 parameters and labels the project MIT, while its Nous-Hermes component is separately licensed under Apache-2.0.
Benchmark scores
| Benchmark | Score |
|---|---|
| AI2D (OpenVLM) | 86.8 |
| Capability | 122.6 |
| HallusionBench (OpenVLM) | 56.5 |
| MathVista (OpenVLM) | 64.3 |
| MM-Vet (OpenVLM) | 61.8 |
| MMMU Validation (OpenVLM) | 55.2 |
| MMStar (OpenVLM) | 64.7 |
| Video-MME | 61.2 |