InternVideo
Shanghai AI LaboratoryInternVideo is Shanghai AI Laboratory's line of video understanding models, built by its OpenGVLab team. InternVideo3, the current release, is an 8B chat model on a Qwen3-VL backbone whose compressed attention cache lets it follow hour-long videos of up to 2,048 frames. While it reasons over a video it can call tools such as temporal grounding, speech recognition and web search, and check its own answer.
The InternVideo2 and InternVideo-Next video encoders in the same repository are vision backbones rather than chat models and are not part of this entry. InternVideo3's weights sit under a team member's Hugging Face account, which the official repository links.
Openness
3 high confidence- weights
- open (InternVideo3-8B-Instruct safetensors on the Hub, ungated)
- data
- partial (the 380K-video long-video SFT set is published, but not the rest of the 7.2M-sample SFT mixture or the continued-pretraining data)
- code
- open (continued-pretraining and SFT configs and training scripts for InternVideo3 in the repository)
- license
- Apache-2.0 (code and weights)
InternVideo3's weights and code are Apache-2.0, and the repository carries its training configs. Only the long-video part of the fine-tuning data is published; the rest of the mixture and the continued pretraining data are described in the report but not released.
- https://huggingface.co/api/models/yanziang/InternVideo3-8B-Instruct?expand[]=downloads&expand[]=gated&expand[]=cardData&expand[]=siblings&expand[]=createdAt recorded 2026-11-12
Hub API for InternVideo3-8B-Instruct: gated false, safetensors shards for the language model, projector and vision encoder.
- https://huggingface.co/yanziang/InternVideo3-8B-Instruct/raw/main/README.md recorded 2026-11-12
InternVideo3-8B-Instruct card: license apache-2.0; long-video understanding up to 2048 frames at 4fps and a 262,144-token context.
- https://raw.githubusercontent.com/OpenGVLab/InternVideo/main/InternVideo3/README.md recorded 2026-11-12
InternVideo3 README: SFT of about 7.2M samples including a 379K-video long-video component; the long-video SFT data is on the Hub as yanziang/InternVideo3_Dataset (380K rows); "This project is released under the Apache 2.0 License."
- https://raw.githubusercontent.com/OpenGVLab/InternVideo/main/LICENSE recorded 2026-11-12
LICENSE: Apache License, Version 2.0, "Copyright (c) OpenGVLab. All rights reserved".
- https://ungh.cc/repos/OpenGVLab/InternVideo/files/main recorded 2026-11-12
Repository tree: InternVideo3/InternVideo3_sft with configs/internvideo3_cpt.py, internvideo3_sft_short.py, internvideo3_sft_long.py and scripts/train.sh.
Adoption
1 medium confidenceHugging Face downloads over the trailing 30 days for InternVideo3-8B-Instruct and InternVideo2.5-Chat-8B. The InternVideo2 encoders and the community quantizations are not counted.
- https://huggingface.co/api/models?author=OpenGVLab&search=InternVideo&limit=100&expand[]=downloads&expand[]=cardData&expand[]=gated&expand[]=createdAt recorded 2026-11-12
InternVideo2_5_Chat_8B: 1,541 downloads in the trailing 30 days.
- https://huggingface.co/api/models/yanziang/InternVideo3-8B-Instruct?expand[]=downloads&expand[]=gated&expand[]=cardData&expand[]=siblings&expand[]=createdAt recorded 2026-11-12
InternVideo3-8B-Instruct: 847 downloads in the trailing 30 days.
Capability
3 medium confidenceInternVideo3 understands images and hour-long video and can call tools while it reasons, level with MiniCPM-V. Its tools inspect the video; they do not operate a live screen or computer.
- https://huggingface.co/yanziang/InternVideo3-8B-Instruct/raw/main/README.md recorded 2026-11-12
Card: "Trained with a short-to-long curriculum (up to 2048 frames at 4fps), supporting hour-long video comprehension"; "Built-in support for recursive perception-action loops with tool use (temporal grounding, ASR, web search, video segmentation)".
- https://raw.githubusercontent.com/OpenGVLab/InternVideo/main/InternVideo3/README.md recorded 2026-11-12
README: video and image understanding examples; evaluation scripts for Video-MME, MLVU, LVBench, MMMU and VSI-Bench.
Verified 2026-11-12