InternVL is a pioneering open-source family of vision-language models from OpenGVLab (Shanghai AI Laboratory), positioned as an open alternative to GPT-4o-class commercial multimodal systems. The repository — home to the InternVL model series, training code, and evaluation tools — has over 10,000 stars on GitHub and powered a CVPR 2024 oral paper, with releases continuing through 2025.
The InternVL approach scales the vision encoder far beyond typical open models — up to 6B and even 22B-parameter ViT variants — and aligns it with strong open-source LLMs, closing much of the gap to proprietary systems on benchmarks for image understanding, OCR, chart and document analysis, and video comprehension. The repo provides model checkpoints, fine-tuning recipes, deployment examples, and a hosted chat demo, covering tasks such as image classification, image-text retrieval, semantic segmentation, and multi-turn visual dialogue. Successive generations — from the original InternVL through InternVL2 and InternVL3 — steadily added native multimodal pretraining, stronger OCR and document skills, and better video support, and the family is widely used as the open backbone inside downstream products and research systems.
Free. InternVL is open source under the MIT license; model weights and code can be downloaded and self-hosted at no cost beyond your own compute.
Researchers and engineering teams who need a strong open vision-language model they can inspect, fine-tune, and deploy themselves — for document AI, visual question answering, multimodal search, and video understanding.