VideoChat3 debuts as fully open 4B video MLLM, 70.1 on Video-MME

A joint team from Nanjing University, Shanghai AI Laboratory, Nanyang Technological University and Peking University has released VideoChat3, a 4B-parameter video multimodal model that ships with weights, training code, the full training recipe and the datasets it was trained on. Most of the open video MLLMs so far have been open in name only, with the alignment data or the recipe held back. This one is the whole stack.

On the numbers the paper reports, the model hits 70.1 on Video-MME without subtitles, 61.7 on MotionBench, 75.6 on TempCompass and 56.7 on LVBench, and the authors say it beats Qwen3-VL-4B on 18 of 19 directly comparable metrics. The architectural bet is what they call an Inflated 3D Vision Transformer, initialised from MoonViT and extended to joint spatiotemporal attention. That gives them a 16x temporal compression ratio on visual tokens, which is where the efficiency claim lives: at 2048 input frames the paper measures a 54% total-latency reduction versus Qwen3-VL at similar quality.

Alongside the model come three datasets that are probably the more durable contribution. VideoChat3-Academic2M packages 2.27M instances of enhanced academic video data, VideoChat3-LV116K adds 116.2K long-form clips with mean durations that stretch up to about 1,300 seconds, and VideoChat3-OL617K covers online streaming with 617K instances built around state tokens for silence, standby and response. The construction pipelines for each are released, so other labs can replay or extend them.

The honest caveat is that these are the authors' own reported numbers on public leaderboards, not independent replications, and the ablation section itself shows the online tuning regresses one streaming subset, ProactiveVQA WEB, by 10.6 points. The paper also notes it surpasses GPT-5 and Gemini 2.5 Flash only on the TimeLens temporal-grounding splits, not on general reasoning, so take it as a strong specialist rather than a universal replacement. What the reporting does not give you is a serving cost picture or a license breakdown for the source clips inside the released datasets.

The bit worth watching is the downstream effect. A fully open 4B model with credible efficiency at long frame counts is exactly the shape a small team or an on-device product wants to fine-tune against, and the released data pipelines make that easier than any prior video MLLM drop this cycle.