How one Korean AI startup's self-improving model just rewrote the benchmark record books
TL;DR: VIDRAFT, a Korean AI startup, announced on September 28, 2026 that its reasoning model Darwin-180B-RSI claimed first place across all five Hugging Face-certified benchmark leaderboards it entered. The model achieved perfect scores on the AIME 2026 and HMMT 2026 mathematics benchmarks — the first time any open model has done so on Hugging Face's official math leaderboards. The results are driven by VIDRAFT's proprietary Recursive Self-Improvement (RSI) technology, which enables the AI to generate and verify its own training data without human intervention.
VIDRAFT, the Korean AI startup behind the Darwin family of reasoning models, made headlines on September 28, 2026, when it confirmed that its latest open model — Darwin-180B-RSI — had simultaneously claimed the top position on five separate Hugging Face-certified global AI benchmark leaderboards, a sweep that the company says no other open model has achieved on these officially sanctioned evaluations.
The star of the announcement is Darwin-180B-RSI, a 180-billion-parameter model built on a Mixture-of-Experts (MoE) architecture that activates only 10 of its 512 expert modules at any given time, keeping inference efficient while preserving massive overall capacity. The model supports a context window of approximately 260,000 tokens.
The five leaderboards on which Darwin-180B-RSI reached first place are all Hugging Face-designated official benchmarks — meaning they undergo a formal vetting process rather than being self-reported or privately administered. The specific results reported by VIDRAFT are:
Notably, this is the first instance of any model achieving a perfect score on Hugging Face's officially certified mathematics leaderboards. On GPQA Diamond, VIDRAFT also holds third place with its earlier Darwin-397B-ZTC model, meaning the company occupies both the first and third positions on that leaderboard simultaneously.
VIDRAFT has released Darwin-180B-RSI as an open model on Hugging Face, where the company currently lists more than 50 official models and over 400 derivative models.
CEO Kim Min-sik stated: "We will continue building AI that keeps growing without human intervention."
The benchmarks in question are not trivial. AIME and HMMT are the qualification and competition rounds used to select national mathematics olympiad representatives and elite university competitors — problems that challenge the best high-school mathematicians in the world. GPQA Diamond problems are authored by domain-specific PhD researchers, and even subject-matter experts in those fields answer correctly only around 65% of the time. MMLU-Pro covers law, medicine, economics, and eleven other disciplines across more than 12,000 questions presented in a ten-choice format.
Sweeping all five with a single open model — particularly achieving perfect scores on the two math competitions — positions VIDRAFT's Darwin series as a serious contender at the frontier of open-model reasoning.
What makes the achievement technically notable is the mechanism behind it. VIDRAFT's Recursive Self-Improvement (RSI) approach differs from conventional knowledge distillation, where a smaller model learns by mimicking the outputs of a larger, more powerful teacher model. Instead, Darwin-180B-RSI solves problems independently, selects only the verified correct solutions from its own attempts, and retrains on that filtered output. The model effectively produces its own high-quality training data in a continuous, compounding loop — without relying on a more capable external model to supply ground truth.
Complementing RSI is VIDRAFT's Zero-Token Confidence (ZTC) technology, which analyzes a model's internal state before it generates a response to estimate the probability of a correct answer in advance. The company describes ZTC as a way to preemptively flag potential hallucinations or agent errors without adding extra inference cost. A "selective merging" technique — described by VIDRAFT as analogous to an MRI-style diagnostic — further distills the strongest components from multiple models and blends them according to reliability weights.
VIDRAFT also drew a direct comparison with Sakana AI, noting that while Sakana AI has pursued a similar self-improvement direction, Darwin's benchmark performance exceeds it, and VIDRAFT's official Hugging Face model count is approximately six times larger, with cumulative downloads more than seven times higher. The underlying technical methodology has been detailed in VIDRAFT's published "Darwin Family" research paper.
Q: What is Darwin-180B-RSI and why did it make news?
A: Darwin-180B-RSI is VIDRAFT's latest open reasoning model, which topped all five Hugging Face-certified AI benchmarks it entered — including the first-ever perfect scores on the AIME 2026 and HMMT 2026 math leaderboards — as announced on September 28, 2026.
Q: What is Recursive Self-Improvement (RSI) and how does VIDRAFT use it?
A: RSI is VIDRAFT's proprietary training methodology in which the AI model solves problems on its own, filters for verified correct solutions, and retrains on those results — generating its own training data without human input or a teacher model, enabling continuous, compounding performance gains.
Q: Is Darwin-180B-RSI publicly available?
A: Yes. VIDRAFT has released Darwin-180B-RSI as an open model on Hugging Face, where it can be downloaded alongside the company's broader library of over 50 official and 400-plus derivative models.
Source: AI타임스 (2026-09-28) — original article