VideoChat3 introduces a fully open-source multimodal large language model (MLLM) designed for efficient and generalizable video understanding across motion analysis, long-form video, and streaming interactions. The model addresses key limitations in existing open-source alternatives, including poor cross-domain generalization,