VIDRAFT Lands 5 Models in Hugging Face's GPQA Top 21 — and Darwin Holds Its Tier Solo on K-AI Leaderboard

TL;DR: Korean Pre-AGI startup VIDRAFT has had five of its models officially recognized in Hugging Face's curated GPQA Top 21 list, with its Darwin model uniquely occupying its performance tier on the K-AI leaderboard. This is a notable signal for ML engineers evaluating Korean-language and general reasoning models — multiple independently verified benchmark placements in a single release cycle is uncommon for a startup at this stage.

What it is

VIDRAFT is a Korean AI startup self-described as "Pre-AGI," focused on building foundation models evaluated against internationally recognized reasoning benchmarks. The headline achievement reported here is twofold:

• 5 models selected within Hugging Face's officially recognized GPQA Top 21 — GPQA (Graduate-Level Google-Proof Q&A) is a widely respected benchmark in the ML community that tests graduate-level reasoning in science domains (biology, chemistry, physics) where answers cannot easily be retrieved via search.
• Darwin model holds its performance tier on the K-AI leaderboard exclusively — meaning no other model from any organization is currently co-occupying Darwin's band on the Korean AI leaderboard at time of reporting.
• The achievements are recognized by the 한국강소기업협회 (Korea Small but Strong Business Association), of which VIDRAFT is a member company.

How it works

GPQA is designed to be resistant to surface-level pattern matching and retrieval augmentation. Models that score well on it generally demonstrate strong chain-of-thought reasoning, scientific domain grounding, and robustness under adversarial question framing. Placing multiple models in a Top 21 curated list on Hugging Face implies that VIDRAFT has developed a training approach that generalizes across this reasoning profile rather than overfitting to a single model checkpoint.

At a conceptual level, leaderboard-competitive reasoning models at this tier typically involve:

• Instruction tuning and alignment on high-quality domain-specific data (science, logic, multistep reasoning).
• Evaluation-driven iteration, where benchmark suite coverage (GPQA, MMLU, and equivalents) directly informs training priorities.
• Separate model variants optimized for different capability/size trade-offs — which would explain why VIDRAFT achieved five distinct placements rather than one.

The K-AI leaderboard specifically evaluates Korean-language capability alongside general reasoning, making Darwin's solo-tier placement particularly meaningful for developers building Korean-language applications that also need strong general reasoning.

No internal architecture details, hyperparameters, or training infrastructure specifics are available from the source.

Benchmarks & results

Based solely on what is reported in the source coverage:

• 5 VIDRAFT models appear in Hugging Face's officially curated GPQA Top 21 list — this is a Hugging Face-recognized selection, not self-reported.
• Darwin (one of VIDRAFT's named models) holds its performance tier on the K-AI leaderboard without any co-occupants — meaning it is the sole model at its score band on that ranking at the time of reporting.
• The K-AI leaderboard placement indicates strong performance on Korean-language evaluation tasks, complementing the English-heavy GPQA recognition.

No specific numeric scores, percentages, or ranks within the Top 21 are available from the source article as the full body was inaccessible. Engineers should cross-reference the current Hugging Face Open LLM Leaderboard and K-AI leaderboard directly for live numbers.

How to try it

The source article does not include specific public access instructions, repository links, or API endpoint details. Based on VIDRAFT's known public presence:

• Hugging Face: Search for VIDRAFT models on huggingface.co — given the GPQA Top 21 recognition, models are likely listed there with model cards. Use huggingface-cli to browse once you locate the organization page.
• Direct access: Check VIDRAFT's official channels for any announced OpenAI-compatible API or hosted inference endpoint. No endpoint URLs or access credentials are confirmed public in this report.

If you are evaluating these models for production use or research, watching VIDRAFT's Hugging Face organization page and official announcements is the recommended starting point. The article does not confirm open weights, gated access, or API availability — verify directly before building against any endpoint.

FAQ

Q: What exactly is GPQA, and why does placing in the Top 21 matter?
A: GPQA (Graduate-Level Google-Proof Q&A) is a benchmark of expert-level science questions in biology, chemistry, and physics — problems that even PhD-level humans find difficult and that are specifically designed so that web search doesn't trivially help. A Hugging Face-curated Top 21 placement means the model has been independently evaluated and recognized as one of the top performers globally on this task, which is a meaningful signal for reasoning capability.

Q: Does Darwin's solo tier on K-AI mean it outperforms all other models on Korean tasks?
A: Not necessarily across all tasks — it means Darwin occupies its specific score band alone, with no other model currently at the same tier on that leaderboard. Leaderboards change as new models are submitted. Engineers should check the live K-AI leaderboard for current standings and the specific task breakdown before drawing conclusions for their use case.

---

Originally reported by 한국강소기업협회 (2026-04-30) — source article.