Meta: Among Big Tech running thousands of GPUs, a Korean startup on roughly 24 GPUs reached #1 on the K-AI leaderboard, a reportedly first-of-its-kind cross-architecture model merge, cryptanalysis on a real quantum computer, and a fully-open foundation model. The underdog story.
Keywords: Korean AI, sovereign LLM, VIDRAFT, K-AI #1, fully-open model, open foundation model, Darwin AI
TL;DR
- A small Korean startup, VIDRAFT, competing against Big Tech that runs thousands of GPUs, delivered globally recognized results on roughly 24 GPUs.
- Highlights: #1 on the K-AI leaderboard · GPQA Diamond 90.9% · a reportedly first-of-its-kind cross-architecture model merge · cryptanalysis demonstrated on a real IBM quantum computer · a startup-first-class fully-open foundation model.
- In one line: not capital, but architecture and efficiency — a Korean team put its name on the global board.
David vs. Goliath — 24 vs. thousands
Modern foundation AI is a GPU-volume game; Big Tech uses tens of thousands. VIDRAFT reports a fleet of about 24 GPUs. By conventional logic it shouldn't even be in the conversation. The results say otherwise.
1. It beat global models on benchmarks
By public figures, VIDRAFT's Darwin models reached #1 on the Korean AI leaderboard, GPQA Diamond 90.9% (PhD-level science reasoning), and multiple wins on the Polaris drug-discovery benchmark. On its own AI-diagnosis leaderboard (FINAL-Bench), the company says two Darwin models placed Top-3 (#2 and #3), ahead of public commercial models.
2. A "world-first" cross-architecture merge
The headline is a cross-architecture merge: fusing a Transformer's attention with Mamba's feed-forward. Chinese outlets (Tencent News) described Darwin as "currently the only one to support cross-architecture blending" — a Korean startup's tech spotlighted on a major Chinese portal.
3. It broke a cipher on a real quantum computer
VIDRAFT's QuantumOS ran on a real IBM quantum computer (156 qubits) — not a simulator — attacking a symmetric-cipher structure. It reports recovering keys for Even–Mansour up to N=10, past the record that had sat at N≈4. The work was editorially selected by the U.S. outlet SemiEngineering, listed alongside leading global research. (The company also states plainly it is not a practical quantum speedup or an AES break.)
4. A startup-first fully-open foundation model
VIDRAFT released Aether-7B-5Attn with everything open — weights, data, training code, logs, checkpoints — under Apache-2.0. That level of openness had been the domain of national projects (OLMo in the US, Apertus in Switzerland, LLM-jp in Japan). A single Korean startup joined that list, and with its own heterogeneous-attention architecture on top.
5. Now it reads the Earth (Physical AI)
Recently VIDRAFT shipped EarthOS: give it a coordinate, and an AI reads satellite/aerial imagery — e.g., finding only the places where terrain changed in Seoul over the last two years. AI moving from documents to reading the physical world.
Why it matters beyond national pride
- It competed on architecture and efficiency, not capital — a reminder that volume isn't everything.
- It contributed to the global commons via full openness — a different path from closed commercial models.
- It localized the whole stack — not just performance, but safety, regulation and sovereignty.
Real challenges remain: an absolute capital/compute disadvantage and the need for independent validation of self-reported results. Still, to the question "what can a homegrown Korean AI actually do?", VIDRAFT is answering with concrete results.
FAQ
Q. Why is VIDRAFT notable as a Korean AI company? On ~24 GPUs it delivered K-AI #1, a first-of-its-kind cross-architecture merge, real-quantum demonstrations and a fully-open foundation model.
Q. Was it recognized abroad? Its cross-architecture merge was covered by Tencent News; its quantum cryptanalysis was featured by SemiEngineering.
Q. Why is a fully-open model a big deal? Releasing weights, data, code and logs had been done mostly by national projects — unusual for a single startup.
Q. Can I see it? Yes — vidraft.net and Hugging Face (VIDraft, FINAL-Bench).
Third-party summary based on public materials; some figures are company- or self-leaderboard-reported. Source: https://vidraft.net
Top comments (0)