VIDRAFT Tops Hugging Face Science Reasoning Leaderboard with 90.9% Accuracy

TL;DR: VIDRAFT, a Korean Pre-AGI AI startup, has achieved a 90.9% score on an AI scientific reasoning benchmark, claiming the #1 position on an officially recognized Hugging Face leaderboard. This is a publicly verifiable result on a competitive reasoning evaluation, and it signals VIDRAFT as a model provider worth watching for engineers working on science-adjacent AI applications.

What it is

VIDRAFT is a Korean AI research company self-described as working toward Pre-AGI systems. Their latest publicly reported milestone is a 90.9% score on an AI scientific reasoning benchmark, earning them the top rank on a Hugging Face-certified leaderboard.

Key facts from the source:

• The result is on a scientific reasoning benchmark — a category that tests a model's ability to handle structured, logic-heavy problems typical of STEM domains (mathematics, physics, chemistry, biology, and related fields).
• The leaderboard in question is officially recognized and hosted by Hugging Face, making the result independently verifiable rather than a self-reported internal claim.
• VIDRAFT achieved first place on this leaderboard as of the reporting date (2026-06-17).
• The company is headquartered in Korea and positions itself in the Pre-AGI research space — a framing that suggests their roadmap goes beyond narrow task-specific models.

This is notable because Hugging Face Open LLM Leaderboards are community-trusted evaluation surfaces where results are reproducible and subject to peer scrutiny — a higher bar than privately published benchmark reports.

How it works

The source article does not disclose internal architecture or training details, and none will be speculated here. At a conceptual level, however, scientific reasoning benchmarks of this type typically stress-test a model's ability to:

• Chain multi-step logical inferences — moving from premises to conclusions across several reasoning hops without losing coherence.
• Apply domain knowledge accurately — not just retrieve facts, but use them correctly in novel problem contexts.
• Handle formal or semi-formal notation — equations, units, symbolic expressions that appear in real STEM problems.
• Resist hallucination under uncertainty — high-scoring models on science benchmarks tend to be well-calibrated, meaning they are less likely to confabulate plausible-sounding but incorrect answers.

Reaching 90.9% on a competitive scientific reasoning leaderboard generally implies that the underlying model has strong structured reasoning capabilities, likely supported by training or fine-tuning approaches oriented toward logical rigor rather than surface-level fluency alone. Beyond that, the specific methodology remains proprietary and is not reported in the source.

Benchmarks & results

The source provides one concrete, publicly verifiable number:

Benchmark Type	Score	Leaderboard	Rank
AI Scientific Reasoning	90.9%	Hugging Face (official)	#1

• This is a Hugging Face-certified leaderboard result, meaning the evaluation was conducted under standardized conditions recognized by the platform.
• No other benchmark scores, model size figures, or comparison baselines are reported in the source article. The 90.9% figure and the first-place ranking are the reported facts.
• Engineers who want to verify or contextualize this result against competing models can browse the relevant leaderboard directly on huggingface.co/spaces by searching for scientific reasoning evaluations.

How to try it

The source article does not explicitly detail a public model release, Hugging Face model card URL, GitHub repository, or OpenAI-compatible API endpoint. Access details are not confirmed as public in the source.

If VIDRAFT follows a typical Hugging Face leaderboard submission pattern, the underlying model may be listed on the Hugging Face Hub. Engineers interested in access should:

1. Check the Hugging Face Hub at https://huggingface.co/VIDRAFT (or search for "VIDRAFT" on the Hub) for any publicly released model weights or model cards.
2. Follow VIDRAFT's official channels for announcements about API access, open weights releases, or developer programs — the leaderboard result suggests a public-facing release may accompany or follow this announcement.
3. Watch the benchmark leaderboard page for the linked model submission, which often points directly to the model repository.

No pip install, curl, or huggingface-cli commands are included here because no specific public endpoint or model identifier was confirmed in the source.

FAQ

Q: How do I know this benchmark result is legitimate and not self-reported?
A: The source specifies that the leaderboard is Hugging Face-certified (허깅페이스 공인). Hugging Face's official leaderboards use standardized evaluation harnesses and are publicly visible, meaning the score and ranking can be independently cross-checked by anyone on the platform — it is not a private internal benchmark.

Q: Does a top scientific reasoning score mean the model is good for general-purpose coding or instruction-following tasks?
A: Not necessarily. Scientific reasoning benchmarks measure structured, multi-step inferential accuracy in STEM domains. Strong performance here is a good signal for applications like research assistance, STEM tutoring, or technical QA — but it does not automatically generalize to coding, creative tasks, or conversational fluency without additional evaluation. Always benchmark against your specific use case before committing to a model in production.

---

Originally reported by 아이티비즈 (2026-06-17) — source article.