VideoChat3 is a fully open 4B-parameter video multimodal large language model from the MCG-NJU lab (Nanjing University's Media Computing Group), built as one generalist model for fine-grained motion understanding, long-video reasoning, precise temporal grounding, and online proactive response to a live stream, rather than separate models per regime. Its stated motivation is that existing open video MLLMs struggle to generalize across such diverse video types while remaining computationally practical, and VideoChat3 addresses both sides at once. On the efficiency side, the model uses an I3D-ViT that compresses redundant spatiotemporal tokens by 16x, combined with an adaptive frame resolution scheme that raises resolution only when a query needs closer inspection of a moment, keeping training and inference cost down as video length grows. On the data side, the team built a synthesis pipeline producing three new training sets released alongside the model: Academic2M for general video understanding, LV116K for long-form video, and OL617K for online, streaming-style video, so the model sees short clips, hour-scale narratives, and live-stream input during training rather than being fine-tuned toward just one afterward. Per the accompanying paper, VideoChat3 reaches this generalist behavior at only 4B parameters, a scale the authors position as competitive with larger open-source alternatives on general, long-form, and streaming video benchmarks. The GitHub release ships every stage needed to reproduce the model, not just the final checkpoint: full-stage training data and per-stage checkpoints, XTuner V1-based training code with staged configs and Slurm/non-Slurm launchers, and an evaluation harness on VLMEvalKit covering both offline and online benchmarks. Two demo scripts ship in the repo -- one for standard single-turn video inference and one for online proactive response, answering during an ongoing stream rather than only after a video finishes. Weights, data, and code were released progressively through July and August 2026, with the paper and evaluation code last updated August 25. The GitHub repository carries no LICENSE file, but the VideoChat3-4B model card on Hugging Face declares an Apache-2.0 license.