Atlas

Models

← All models

InternVL3-8B

OpenGVLab · 2025-04-11 · 7.9B parameters

InternVL3-8B is OpenGVLab's 7.94B-parameter multimodal checkpoint, combining an InternViT-300M vision encoder, a two-layer MLP projector, and a Qwen2.5-7B language backbone. Released on April 11, 2025 after native multimodal pre-training, supervised fine-tuning, and mixed preference optimization, it accepts text, image, and video within a 32,768-token context; its weights are Apache-2.0 licensed while the surrounding InternVL project code is MIT licensed.

Benchmark scores

BenchmarkScore
AI2D (OpenVLM)85.1
Capability123.7
CharXiv Reasoning37.6
CharXiv-D73.6
HallusionBench (OpenVLM)49.0
MATH-Vision29.0
MathVista (OpenVLM)70.5
MM-Vet (OpenVLM)82.8
MMMU Validation (OpenVLM)62.2
MMStar (OpenVLM)68.7
Loading Atlas data…