VIDRAFT Hits 90.9% on GPQA Diamond, Pushing the Bar for AI Scientific Reasoning
TL;DR: VIDRAFT, a Korean Pre-AGI AI startup, has reported a 90.9% score on the GPQA Diamond benchmark — one of the most demanding evaluations of graduate-level scientific reasoning for AI systems. This result positions VIDRAFT among a very short list of models achieving frontier-level performance on expert science questions. Developers working on reasoning-heavy applications in chemistry, biology, and physics should take note.
What it is
VIDRAFT is a Korean AI company focused on building Pre-AGI systems, with a stated emphasis on deep scientific reasoning capabilities. Their latest reported achievement centers on GPQA Diamond, a benchmark designed to test AI on graduate-level multiple-choice questions authored by domain experts in fields such as biology, chemistry, and physics — questions difficult enough that even PhD-level humans outside the relevant specialty frequently answer incorrectly.
The headline result is a 90.9% accuracy on GPQA Diamond, which the company describes as evidence that their approach to AI scientific reasoning is technically validated and competitive at a global frontier level.
This is not a general-purpose chatbot milestone. GPQA Diamond is specifically constructed to resist surface-level pattern matching: questions require genuine multi-step reasoning chains grounded in expert scientific knowledge. Clearing 90% on this benchmark is, as of mid-2026, a mark achieved by only a handful of the world's most capable large reasoning models.
How it works
While VIDRAFT has not publicly disclosed the full architectural or training details of the underlying model, the GPQA Diamond result gives developers useful conceptual signal about the system's design priorities:
- Chain-of-thought and multi-step inference: Achieving high accuracy on GPQA Diamond almost universally requires a model capable of decomposing complex scientific problems into structured reasoning steps rather than retrieving cached answers.
- Scientific domain depth: The benchmark spans biology, chemistry, and physics at a graduate level, meaning the model must generalize scientific reasoning across disciplines rather than narrowly memorizing domain facts.
- Robustness to distractor options: GPQA is specifically designed with plausible-but-wrong answer choices crafted by experts, which means performance here reflects genuine reasoning quality rather than elimination heuristics.
VIDRAFT frames this result as a validation of their AI scientific reasoning technology — suggesting their research direction prioritizes structured, verifiable inference over generative fluency alone. This aligns with a broader trend in frontier model development toward "thinking" or "reasoning" model architectures that dedicate additional compute at inference time to problem-solving.
Benchmarks & results
The sole benchmark figure reported in the source article is:
| Benchmark | Score |
|---|---|
| GPQA Diamond | 90.9% |
Context for developers:
- GPQA Diamond is widely regarded as one of the hardest publicly available science reasoning benchmarks. Human expert performance on out-of-domain questions typically falls in the 65–70% range, making 90%+ a meaningful threshold.
- As of the article's publication date (June 17, 2026), this score places VIDRAFT's model in contention with the top tier of publicly evaluated frontier systems.
- No additional benchmark figures (MMLU, MATH, HumanEval, etc.) were included in the source reporting.
How to try it
The source article does not announce a public model release, Hugging Face repository, GitHub codebase, or developer API endpoint at this time. VIDRAFT has not, based on available reporting, made this specific model publicly accessible through an OpenAI-compatible API or open-weight download.
If you want to stay informed about access:
- Watch VIDRAFT's official channels for announcements of public model releases or API access programs.
- Check Hugging Face (
huggingface.co) and GitHub for any VIDRAFT-affiliated repositories as they become available.
This article will be updated if public access channels are announced.
FAQ
Q: How does GPQA Diamond differ from benchmarks like MMLU, and why does the distinction matter for evaluating reasoning models?
A: MMLU covers a broad range of academic subjects at varying difficulty levels and is susceptible to models performing well through memorization of common facts. GPQA Diamond is specifically designed by domain PhD experts to include only questions that stump non-specialist PhD holders — meaning wrong answers are deliberately plausible. A model scoring 90%+ on GPQA Diamond is demonstrating genuine multi-step scientific reasoning, not retrieval or pattern-matching on commonly seen training data. For developers building applications in scientific research assistance, drug discovery tooling, or technical Q&A, GPQA Diamond performance is a more meaningful signal than MMLU alone.
Q: Is this an open-weight model, and can I run it locally or via API today?
A: Based on current reporting, the model is not publicly released as open weights, and no developer API has been announced. VIDRAFT has shared the benchmark result as a demonstration of their technical capability, but access details have not been made public. Developers interested in early access or research collaboration should monitor VIDRAFT's official communications directly.
Originally reported by AI타임스 (2026-06-17) — source article.
Top comments (0)