InternVL 2.5

OpenGVLab's InternVL line couples an InternViT image encoder with an existing language model, Qwen2.5 or InternLM2.5 depending on size, through an MLP projector. Version 2.5, released in December 2024, spans seven checkpoints from 1B to 78B parameters; smaller ones use a 300M vision tower, while 26B and up use the 6B InternViT. Large images are split into 448 pixel tiles and compressed with pixel unshuffle, keeping token counts manageable for dense documents, charts, screenshots, multi-image prompts and sampled video frames. Beyond inference code, the repo carries fine-tuning recipes including LoRA, the MPO preference-optimization method and its data pipeline, evaluation scripts, format converters and a Streamlit chat demo, with LMDeploy and Transformers as the usual serving routes. InternVL3 and 3.5 live in the same repo. A single consumer GPU handles the 1B to 8B models; 38B and 78B need multi-GPU nodes. Code and 2.5 weights are MIT, but each checkpoint also inherits its bundled LLM's license, so check Qwen terms before commercial use. Aimed at researchers and teams building visual document understanding.