VIDRAFT Tops Hugging Face Leaderboard with 90.9% on GPQA Diamond — Here's What Engineers Need to Know
TL;DR: Korean Pre-AGI AI startup VIDRAFT has recorded a 90.9% score on the GPQA Diamond benchmark, claiming the #1 position on the Hugging Face official leaderboard. GPQA Diamond is one of the most demanding reasoning benchmarks available, requiring graduate-level expertise in science and STEM disciplines. If you track frontier model capabilities, this result puts VIDRAFT firmly on the map.
What it is
VIDRAFT is a Korean AI startup positioning itself in the Pre-AGI space — meaning its research and engineering efforts are aimed at building systems that approach, but have not yet claimed to reach, artificial general intelligence. The milestone reported here is a leaderboard-verified score of 90.9% on GPQA Diamond, as recognized officially by Hugging Face.
GPQA Diamond (Graduate-Level Google-Proof Q&A) is a rigorously curated benchmark consisting of multiple-choice questions written and validated by domain experts — primarily in biology, chemistry, and physics. Questions are deliberately designed to be "Google-proof": answerable only through genuine expert-level reasoning, not surface-level retrieval or pattern-matching. Human experts with PhDs in the relevant field typically score in the range of 65–70% on this benchmark, making a score approaching 91% a significant signal about a model's deep reasoning capability.
The Hugging Face Open LLM Leaderboard is a widely respected, community-monitored ranking that uses standardized, reproducible evaluation conditions. A #1 ranking there carries credibility precisely because submissions are independently verified rather than self-reported.
How it works
VIDRAFT has not published a full technical report at the time of this article, so specific architectural details remain undisclosed. At a conceptual level, however, what is publicly known aligns with current frontier approaches to high-stakes reasoning benchmarks:
- Advanced reasoning pipelines: State-of-the-art performance on GPQA Diamond generally involves models that can decompose multi-step scientific problems, apply domain knowledge, and self-verify intermediate reasoning steps — often through chain-of-thought or extended thinking mechanisms.
- High-quality training data curation: Achieving graduate-level accuracy in STEM domains typically requires deliberate investment in the quality and domain coverage of training data, rather than simply scaling raw token counts.
- Rigorous evaluation alignment: The fact that this result is certified on Hugging Face's official leaderboard — rather than reported in isolation — suggests that VIDRAFT's evaluation pipeline is reproducible and standardized, which matters for comparability.
Beyond that, the internal mechanisms, training configurations, and infrastructure specifics are not publicly disclosed, and this article will not speculate on them.
Benchmarks & results
The single headline result reported from this source:
| Benchmark | Score | Leaderboard Position |
|---|---|---|
| GPQA Diamond | 90.9% | #1 (Hugging Face Official Leaderboard) |
A few points of context that help engineers interpret this number:
- GPQA Diamond is widely regarded as one of the hardest publicly available LLM benchmarks, specifically because it tests for reasoning that cannot be easily gamed through memorization or retrieval.
- A human expert baseline on GPQA Diamond sits around 65–70% — VIDRAFT's reported score significantly exceeds that.
- No other benchmark numbers are cited in the source article at this time. Additional evaluation results may follow in a future technical report or model card.
How to try it
At the time of writing, VIDRAFT has not published specific public access instructions — such as a Hugging Face model repository, GitHub codebase, or OpenAI-compatible API endpoint — in the source article covered here.
Given that the result is verified on the Hugging Face Official Leaderboard, the model or evaluation artifacts may become accessible through Hugging Face channels. Developers interested in access should:
- Watch VIDRAFT's Hugging Face profile for any public model or dataset releases.
- Follow VIDRAFT's official channels for announcements about API access or open-weight releases.
- Check the Hugging Face Open LLM Leaderboard submission details, which may link to a model card or repository once public.
This article will not fabricate endpoint URLs, model names, or curl commands that are not confirmed public.
FAQ
Q: How does 90.9% on GPQA Diamond compare to other well-known frontier models?
A: GPQA Diamond scores for leading models have been climbing steeply through 2024–2025. A score of 90.9% placing #1 on the Hugging Face leaderboard indicates it surpasses all other currently ranked submissions at the time of this report. For precise head-to-head comparisons, consult the live Hugging Face Open LLM Leaderboard directly, as rankings update continuously.
Q: Is this an open-weight model, or is access restricted?
A: The source article does not specify whether the model weights are open or access is API-only. No public download or API endpoint has been confirmed at this time. Monitor VIDRAFT's official announcements for access details.
Q: Why does the Hugging Face leaderboard ranking matter more than a self-reported score?
A: The Hugging Face Open LLM Leaderboard runs evaluations under standardized, independently reproducible conditions. This reduces the risk of benchmark contamination, cherry-picked prompting strategies, or evaluation protocol discrepancies that can inflate self-reported numbers — making the certification meaningful for practitioners comparing models.
Originally reported by 비하인드 (2026-06-18) — source article.
Top comments (0)