German non-profit group LAION, together with the Tübingen AI Center, University of Tübingen, and Forschungszentrum Jülich, released the open video dataset LAION-BVD: 80 million downloaded videos totaling 10 million hours, cut into 55 million clips, plus 300 million frames and two audio subsets of 1.7 million and 10 million tracks. Lightweight versions with URLs and metadata are already freely available on Hugging Face, while full versions with media files are provided upon request and for research purposes only. According to the arXiv:2608.24845 paper, ViCLIP and CLAP models trained on BVD are comparable to those trained on InternVid and outperform them by up to +2.1% on video-text benchmarks.

image
image

What happened

LAION-BVD (Big Video Dataset) was collected from 1.3 billion links in CommonCrawl archives, from which 80 million videos totaling 10 million hours were successfully downloaded. Content-aware scene change detection cut the material into 55 million clips (BVD-V-55M sample), to which synthetic video and audio descriptions were built. Separately, 300 million frames (BVD-I-300M) and two audio subsets of 1.7 million and 10 million tracks (BVD-A-1.7M and BVD-A-10M) were extracted. The release is described in the arXiv:2608.24845 paper dated August 25, 2026. Lightweight versions of all subsamples — only URLs and metadata, including BVD-URLs with 1.35 billion records and BVD-V-55M-URLs — are open on Hugging Face, while full versions with media files are provided upon request.

Context

Until now, corpora of this scale for multimodal pre-training have been concentrated in proprietary companies, and the open scientific community had to work with significantly smaller data. BVD is the first to open a 10-million-hour corpus in three modalities: video-text for ViCLIP, audio-text for CLAP, and image-text for CLIP-style models. Frames from video deserve special attention: they carry a distribution of real angles, motion, and compositions that is not present in photo banks. The scientific signal of the release is also substantial: ViCLIP and CLAP models trained on BVD outperform those trained on InternVid by up to +2.1% on video-text benchmarks, and when scaling from 10 million to 50 million clips, quality grows noticeably — meaning the "data vs. quality" curve has not saturated, and it can now be measured in an open cycle just like corporate ones.

Why this matters for the industry

For the industry, the main effect is the reduction of entry costs into multimodal video: a big tech-scale corpus is now available not only to closed teams, so the R&D phase of video-text and audio-text search, audio categorization, and agentic "viewing" pipelines no longer requires proprietary data. The open scaling curve allows teams to measure the return from data on their own evals and plan pre-training based on a reproducible dependence on the number of clips. It is expected that in the coming months, independent reproductions of the comparison with InternVid, open checkpoints of ViCLIP and CLAP trained on BVD, derivative works on subsamples, and filtering and deduplication tools around the URL corpus will appear. At the same time, the research-only license makes the dataset a prototyping outline for startups: a product release based on it will require either a different license or its own data extraction, and the demand for licensed commercial corpora will likely grow under such circumstances.

Why this matters for users

If you want to train or fine-tune your video or audio model, you can try it today: download the URL versions of the subsamples from Hugging Face, filter the corpus for your task, build a pilot fine-tuning in the style of ViCLIP or CLAP on a 10-million-clip subsample, and compare the result with your current baseline model on your own evals — this is a hypothesis check without the cost of closed data. In parallel, you can apply for the full versions with media files through the project questionnaire. Two practical limitations: the license allows only research use, so a commercial product cannot be built on this data, and working with the URL version means your own stage of downloading videos by external links with all the consequences for broken addresses.

What is still unknown / limitations

The increase of up to +2.1% over InternVid without confidence intervals and independent replication should be correctly read as parity with a slight advantage, not as a gap. The URL version is not equal to the corpus: between the metadata and training lies the stage of downloading 80 million videos by external links, which is not reproducible by default without a detailed description. Full media files are provided upon request, not freely, and the license is strictly research-only. It is worth checking the long-term stability of the resource: who and how will support such a scale of corpus, it does not follow from the paper.

Sources

Author

Look at AI, editorial team