Paper summary
Paper summary: InternVL3: Exploring Advanced Training and Test-Time Recipes for Open-Source Multimodal Models
InternVL3 trains multimodal and language abilities together in one pre-training stage, rather than adapting a text-only LLM afterwards, and reports 72.2 on MMMU for its 78B model.
- Native multimodal pre-training learns vision and language jointly from the start, instead of bolting vision onto a finished LLM.
- InternVL3-78B scores 72.2 on MMMU, a new open-source state of the art at the time.