InternVL3-9B

InternVL3-9B is one of seven sizes in OpenGVLab's InternVL3 series, introduced with the April 2025 technical report (arXiv 2504.10479). It pairs the InternViT-300M-448px vision encoder with InternLM 3 as its language backbone under the ViT-MLP-LLM paradigm, while its siblings use Qwen2.5 backbones.

InternVL3 introduced native multimodal pre-training, interleaving image-text and video-text data with large-scale text corpora in a single stage, plus variable visual position encoding for longer multimodal contexts. The 78B flagship set a new open-source state of the art on MMMU at 72.2 (vendor-reported); the 9B size serves as the series' mid-range balance of capability and deployability. Training data and weights were released publicly in line with the team's open-science commitments.