On 25 August 2026, LAION released a paper and nine Hugging Face repositories for LAION-BVD (Big Video Dataset): 1.3 billion video URLs, 80 million downloaded videos, ten million hours of footage.[1][2] The abstract calls it an unprecedented expansion of open access to multimodal video.[1] I went to fetch the actual release, and the gap between the headline and the download button is worth a close look — the numbers are real, the openness is layered.
The source corpus is 1.3 billion platform-specific video URLs collected from CommonCrawl; from those, the team downloaded 80 million videos totalling about 10 million hours.[1][2] The platforms named in the card are YouTube, Vimeo and Dailymotion.[2] Content-aware scene detection split the footage into clips, and synthetic captions were generated for both video and audio.[1] The release is organized in layers, each with its own repository:
Most of the clip-level assets also have -URLs companion cards that ship the metadata without the media.[2]
The word "open" carries different weight at each layer. The URL/metadata cards are openly downloadable — the main URL card is CC-BY-4.0 and contains no video content, only URLs and crawl metadata.[2] But the clip subset is gated: "Access to dataset laion/BVD-V-55M is restricted. You must have access to it and be authenticated to access it." — that is the current text of its README, and its gating mode is manual approval.[3]
The raw corpus — the actual videos, the 10 million hours in the title — is not on Hugging Face at all. The card states it plainly: "BVD-RAW itself — the raw pool of 80M videos totalling 10M hours — is not distributed through Hugging Face," and that it and the gated subsets are "available to academic and non-commercial researchers through a single central access request form."[2] So the open release is: all the fingerprints (URLs, crawl metadata, clip metadata and synthetic captions) plus a 55M-clip subset behind manual approval, while the footage itself travels by request to academic and non-commercial researchers only.
The project page calls LAION-BVD "the largest openly accessible video corpus for multimodal learning research."[4] That is true in a meaningful sense — URL-level access at this scale is new — but it is not the same as torrent-and-go, the way CommonCrawl or LAION's image sets have historically been consumed.
Video is the modality where open data is thinnest, and 10 million hours is an order of magnitude beyond what the field has had. The paper reports competitive performance on video-text and audio-text benchmarks with improvements scaling with model size, and it explores video frames as a distributionally distinct image-text source.[1] If the access model holds — URL cards open, raw media request-gated — it creates a two-speed research economy: anyone can build on the metadata, but the pixels flow mainly to academic labs.
Two questions the pages I read do not answer, flagged rather than smoothed over: the permission/consent status of the source platforms (the cards name YouTube, Vimeo, Dailymotion but state no platform agreements),[2] and the platform share of the 1.3B URLs — "e.g. YouTube, Vimeo, Dailymotion" is the extent of the breakdown in the card.[2] LAION's web-scale scraping has a contested legal history; how BVD's platform sourcing is squared with it is not stated in the release materials I could fetch.
[1] arxiv.org/abs/2608.24845 — LAION-BVD: A 10-Million-Hour Open Video Dataset for Multimodal Pre-training (submitted 2026-08-25)
[2] huggingface.co/datasets/laion/BVD-URLs — LAION-BVD 1.3B Video URLs card (verified live 2026-08-26)
[3] huggingface.co/datasets/laion/BVD-V-55M/raw/main/README.md — clip subset README, currently access-restricted (verified live 2026-08-26)
[4] projects.laion.ai/bvd — LAION-BVD project page (verified live 2026-08-26)