The Korean AI startup's latest reasoning model sets new records in math, science, and multimodal evaluation.
TL;DR: VIDRAFT's Darwin-180B-RSI reasoning model claimed first place across five Hugging Face-certified leaderboards, as announced on September 28, 2026. The model achieved perfect scores on the AIME 2026 and HMMT 2026 mathematics benchmarks — the first time any model has done so on those official leaderboards. It also posted a 94.44% score on GPQA Diamond, surpassing previous top-ranked competitors.
VIDRAFT, the Korean AI startup developing pre-AGI systems, announced on September 28, 2026 that its latest reasoning model, Darwin-180B-RSI, has reached first place on five separate Hugging Face-certified benchmark leaderboards, spanning advanced mathematics, graduate-level science, and multimodal reasoning tasks.
Darwin-180B-RSI earned top rankings across all five Hugging Face leaderboards on which it was evaluated. The most striking results came from two olympiad-level mathematics benchmarks: AIME 2026 and HMMT 2026, where the model recorded 100% scores — a feat that, according to VIDRAFT, had never been achieved on those certified leaderboards before. Previous record-high scores on those two benchmarks stood at 97.1% and 92.7%, respectively.
On GPQA Diamond — a benchmark comprising science questions authored by PhD-level domain experts — Darwin-180B-RSI scored 94.44%, edging out Kimi-K3, which had previously led with a score of 93.5%. The model also posted strong results on MMLU-Pro, covering 14 academic disciplines including law, medicine, and economics, and on MMMU-Pro, a college-level multimodal benchmark that requires interpreting visual inputs such as charts, sheet music, and medical imagery.
VIDRAFT attributed the results to three proprietary technologies working in combination. The first is Darwin, a selective model-merging technique that analyzes performance at the level of individual layers and expert units across multiple models, then identifies and combines each model's strongest components. VIDRAFT noted that this approach has been scaled to support models with hundreds of billions of parameters.
The second technology, Recursive Self-Improvement (RSI), enables the model to solve problems independently, verify its own answers, and then retrain on the verified reasoning outputs. The improved model then becomes the solver for the next training iteration — a self-reinforcing cycle designed to raise performance without human-labeled data. VIDRAFT reported that in this model, RSI reduced reasoning chain length by 11% while maintaining the same level of accuracy.
The third technology, Zero-Token Confidence (ZTC), allows the model to assess its own confidence in a potential answer before generating a response. If the model's internal confidence falls below a threshold, it is designed to withhold the answer rather than produce an unreliable output.
VIDRAFT CEO Kim Min-sik commented that the results demonstrate the viability of AI systems that learn autonomously and are self-aware of their own confidence levels, and expressed the company's ambition to build AI that continues to improve without human intervention.
Benchmark performance at this level carries meaningful weight in the AI industry. GPQA Diamond is widely regarded as one of the most demanding evaluation sets for scientific reasoning, given that its questions are designed to challenge even the experts who created them. Achieving a perfect score on olympiad-caliber math benchmarks like AIME and HMMT signals a qualitative leap in structured problem-solving, an area that has long been considered a frontier challenge for large language models.
What makes VIDRAFT's results particularly notable is the breadth of coverage: top rankings across mathematics, expert-level science, multi-domain knowledge, and multimodal understanding suggest that Darwin-180B-RSI is not a model optimized narrowly for one task type. For the broader AI ecosystem, independent Hugging Face certification of these scores lends external credibility that internal evaluations alone cannot provide.
The achievement also underscores the rising competitiveness of Korean AI startups on the global stage. VIDRAFT has now established a track record on publicly audited leaderboards, positioning itself alongside well-resourced international labs in the reasoning model race.
Q: What is VIDRAFT's Darwin-180B-RSI?
A: Darwin-180B-RSI is VIDRAFT's latest large-scale reasoning model, designed for high-difficulty tasks in mathematics, science, and multimodal understanding. It incorporates three proprietary technologies: Darwin selective merging, Recursive Self-Improvement, and Zero-Token Confidence.
Q: Which benchmarks did Darwin-180B-RSI top, and what scores did it achieve?
A: The model ranked first on five Hugging Face-certified leaderboards: AIME 2026 (100%), HMMT 2026 (100%), GPQA Diamond (94.44%), MMLU-Pro (88.12%), and MMMU-Pro (79.48%).
Q: What makes VIDRAFT's Zero-Token Confidence (ZTC) technology distinctive?
A: ZTC enables the model to evaluate its own confidence in a response before generating it. If confidence is assessed as insufficient, the model is designed to withhold its answer, reducing the risk of producing unreliable outputs.
Source: 지디넷코리아 (2026-09-28) — original article