The Korean AI startup's self-improving model outscores Google, DeepSeek, and xAI on a major benchmark.
TL;DR: VIDRAFT, a Korean AI startup, announced on November 6, 2026 that its Darwin-180B-RSI-R3 model achieved the second-highest score in the math and science reasoning category on the SuperCLUE November 2026 evaluation, trailing only OpenAI's GPT-6.1 Sol among 19 competing models. Darwin outperformed flagship models from Google, DeepSeek, and xAI, and ranked first among all open-weight models in that category. The result was driven by VIDRAFT's proprietary Recursive Self-Improvement (RSI) technology, which enables the model to enhance its own reasoning capabilities without requiring new human-labeled data.
VIDRAFT, the Korean AI startup behind the self-improving Darwin model series, made a significant splash in global AI benchmarking on November 6, 2026, when it revealed that Darwin-180B-RSI-R3 had claimed the second-highest math and science reasoning score on the SuperCLUE November 2026 leaderboard — placing ahead of well-resourced competitors from the United States and China alike.
According to AI타임스, VIDRAFT CEO Kim Min-sik confirmed that Darwin-180B-RSI-R3 scored 73.68 points in the math and science reasoning category of the SuperCLUE November 2026 evaluation. That placed the model second out of 19 total models assessed, behind only OpenAI's GPT-6.1 Sol, which recorded 76.32 points.
The Darwin model beat out notable competition: Google's Gemini 4 Argon and DeepSeek's V4.1 Flash both scored 71.05 points, while xAI's Grok 4.7 came in at 68.42 points. Among open-weight models — those with publicly released weights — Darwin ranked first in the math and science reasoning category.
SuperCLUE is a Chinese evaluation body that benchmarks large language models across five domains: agentic coding, math and science reasoning, hallucination suppression, precise instruction following, and long-context processing. The November 2026 round included 19 major AI models from the United States and China; Darwin was the only Korean model in the field.
In the five-domain composite score, Darwin recorded 43.08 points overall, placing ninth among all models and fifth among open-weight models. The company acknowledged that while Darwin excels in math and science reasoning, it still lags behind some competitors in areas such as long-context handling and instruction following.
VIDRAFT also highlighted Darwin's performance on the Hugging Face GPQA Diamond benchmark — a PhD-level science reasoning test covering physics, chemistry, and biology — where the Darwin model family scored 94.44 points as of its November 3 publication. Additionally, the company reported improvements on the general knowledge benchmark MMLU-Pro, alongside an 11% reduction in the reasoning token length required to generate answers.
The technical foundation of Darwin's performance is VIDRAFT's Recursive Self-Improvement (RSI) technology. Darwin-180B-RSI-R3 was built on Alibaba's Qwen3.8-Flash-Next as a base model. Rather than retraining the entire network, VIDRAFT preserved the expert layers responsible for knowledge storage while applying RSI to iteratively improve the reasoning-related weights. The model autonomously generates problems, works through solutions, and verifies its own outputs — repeating this loop without any new human-annotated data being introduced.
The impact is quantifiable: the base Qwen3.8-Flash model scored 57.89 points in SuperCLUE's math and science reasoning category, while Darwin's RSI process pushed that figure up to 73.68 — a gain of 15.79 points. The composite SuperCLUE score also rose from 39.81 to 43.08.
Darwin uses a Mixture-of-Experts (MoE) architecture with 180 billion total parameters, activating approximately 3 billion during inference. Its weights are publicly available, enabling commercial use. VIDRAFT also offers a compressed variant called POCKET-Darwin-180B, designed to run on personal computers and laptops without a dedicated GPU. Cumulative downloads of the Darwin model family on Hugging Face have surpassed 4.68 million.
VIDRAFT received GPU resource support through a project selected by the National IT Industry Promotion Agency (NIPA), which operates under South Korea's Ministry of Science and ICT.
Darwin's SuperCLUE result is a meaningful signal for the global AI landscape. It demonstrates that a Korean startup operating with more constrained resources than Big Tech can still compete at the frontier of reasoning benchmarks — particularly through architectural efficiency and self-improvement techniques rather than brute-force scaling. The RSI approach is especially noteworthy: if models can reliably improve their own reasoning without continuous human data labeling, it could represent a meaningful shift in how AI development scales. Darwin's open-weight release also broadens access, allowing developers and researchers worldwide to build on its capabilities.
Q: How did VIDRAFT's Darwin model perform on the SuperCLUE November 2026 benchmark?
A: Darwin-180B-RSI-R3 scored 73.68 points in the math and science reasoning category, placing second out of 19 models — behind OpenAI's GPT-6.1 Sol — and first among all open-weight models. Its overall composite score was 43.08, ranking ninth overall and fifth among open-weight models.
Q: What is Recursive Self-Improvement (RSI), and why does it matter for Darwin?
A: RSI is VIDRAFT's proprietary technique that enables the Darwin model to generate its own problems, solve them, and verify the results in an iterative loop — improving its reasoning capabilities without needing new human-annotated data. This process raised Darwin's math and science reasoning score by 15.79 points above the base model it was built on.
Q: Is the Darwin model publicly available?
A: Yes. VIDRAFT has released Darwin's model weights as open-weight, permitting commercial use. A compressed version called POCKET-Darwin-180B is also available for users who want to run the model on personal computers or laptops without a dedicated GPU.
Source: AI타임스 (2026-11-06) — original article