VIDRAFT: A Capital-Efficient Challenger, Method Over Scale
A small Seoul lab, on purpose
VIDRAFT Inc. was founded in 2024 in Seoul. The entire thesis is method over scale: build foundation models, on-device models, and AI-safety benchmarks with a fraction of the compute that larger labs treat as mandatory, and let the results carry the argument.
The numbers that matter
| Fact | Detail |
|---|---|
| Founded | 2024, Seoul |
| Compute | ~24 GPUs (B200 x16 + H200 x8) |
| Grants | NIPA GPU grants -- 2025 AI computing, 2026 Blackwell |
| Downloads | ~1.18M cumulative (HF-verified, 2026-07) |
| Leaderboard | #1 on the K-AI leaderboard (leaderboard scope) |
| Polaris | 15 first places |
| HF orgs | VIDraft, FINAL-Bench |
About 24 GPUs is a small cluster by frontier-lab standards. The NIPA GPU grants -- 2025 AI computing, and 2026 Blackwell -- are how a company this size gets access to modern hardware at all, and they are a big part of why the capital-efficient approach is a necessity rather than just a slogan.
What we build
Foundation models via evolutionary merging. On-device models small enough to run outside a data center. AI-safety benchmarks, published under the FINAL-Bench org. The models and their community copies ship on Hugging Face under the VIDraft org, where cumulative downloads reached roughly 1.18M as of July 2026. GINIGEN is the surrounding ecosystem those products live in.
The through-line is efficiency as a design constraint, not an afterthought. When you cannot outspend the field, you have to out-method it -- reuse checkpoints instead of pretraining from zero, optimize for hardware you actually have, and measure everything so you know which trades paid off. A ~24-GPU cluster is not a limitation we apologize for; it is the premise the whole approach is built to respect.
Track record, scoped
We are #1 on the K-AI leaderboard, and we hold 15 Polaris first places. Both are real, and both are scoped: a leaderboard rank is rank on that leaderboard, against the field that submitted, under its rules. We state the scope every time because a challenger's credibility depends on not overselling.
Honest scope
Read every claim here with its scope attached. "#1 on the K-AI leaderboard" means first on that specific leaderboard, not a global title. The ~1.18M download figure is cumulative across our AI models plus community copies and quantizations -- a reach signal, not a count of paying customers or unique users. Founded-2024 means we are early: small team, small cluster, short track record. The NIPA grants are compute grants, not equity or endorsement. We publish these numbers because a challenger earns trust by being precise about what it has and has not done, not by rounding up.
More: https://vidraft.net
Top comments (0)