InternVL3-78B

An update to the InternVL series, which uses various sized InternViT encoders and different versions of Qwen2.5 (and InternLM) for the various model sizes. In this update, they introduce multimodal pre-training, i.e., using interleaved text-image data during the pre-training stage, followed by a SFT and an RL phase. The benchmark scores of the model are very good.