Article: VIDRAFT’s 398-billion-parameter model, Darwin-398B-JGOS, clinched third place on the GPQA Diamond leaderboard using just 24 GPUs. The result proves a small Korean startup can rival the world’s biggest AI labs on a test that demands genuine scientific reasoning instead of memorized answers.

Why GPQA Diamond matters

GPQA Diamond asks graduate-level science questions that don’t appear on public web sites. Because the answers can’t be scraped, a model must reason through the problem rather than retrieve a stored fact. The benchmark forces logic, not memorization.

The current leaderboard

  • Kimi-K3 (CN): 93.5
  • GLM-5.2 (CN): 91.2
  • Darwin-398B-JGOS (KR): 90.9
  • DeepSeek-V4-Pro (CN): 90.1
  • Inkling-Small (US): 89.5
  • Qwen3.5-397B-A17B (CN): 88.4
  • Nemotron-3-Ultra-550B (US): 87.9

Darwin outruns models from Nvidia, DeepSeek and several US and Chinese offerings, even though it runs on a fraction of the hardware typical for such projects.

The method behind the win

VIDRAFT didn’t buy a massive GPU farm. Instead, the team used “evolutionary merging,” iteratively combining multiple open-source models and keeping the best-performing components. Running this on a modest 24-GPU cluster yielded a high-quality, 398-billion-parameter network without the capital outlay usually needed to train from scratch.

What this means for the AI field

  • Lowered entry barrier: Building elite AI no longer requires thousands of GPUs.

Caveats and counter-points

Darwin still trails the two leaders by a narrow margin, and its 398-billion-parameter size remains massive—training or serving the model still demands significant infrastructure. The approach hinges on evolutionary merging of open models on a small cluster. GPQA Diamond served as the benchmark.

What to watch next

The takeaway is clear: clever model-fusion strategies can offset raw compute, reshaping who can compete at the highest levels of AI research.