The Korean AI startup's self-improving model outscores Gemini and DeepSeek on a major benchmark — while making its weights freely available.
TL;DR: VIDRAFT, a Korean AI startup, publicly released the weights for Darwin-180B-RSI-R3 on November 6, 2026. According to the company's published figures, the model scored 73.68 on SuperCLUE's November 2026 math and science reasoning evaluation, ranking first among all open-weight models in that category. It placed second overall in the category, behind only OpenAI's closed GPT-6.1-Sol (76.32).
VIDRAFT made headlines in the global AI community on November 6, 2026, when the Korean AI startup published the full weights of Darwin-180B-RSI-R3 — its latest recursively self-improved language model — to Hugging Face. The release was covered by Japanese AI media outlet 라보메모 ラボメモ as part of its November 7, 2026 daily news roundup, which highlighted the model's standout performance on a Chinese third-party benchmark for mathematical and scientific reasoning.
VIDRAFT released the model weights for Darwin-180B-RSI-R3, a large language model built through a process called recursive self-improvement (RSI). According to figures published by the company and cited in 라보메모's report, the model was evaluated by SuperCLUE — a Chinese AI evaluation body — in its November 2026 assessment using a non-public test set.
On the math and science reasoning sub-index, Darwin-180B-RSI-R3 scored 73.68. That placed it second overall among the 19 models evaluated in that category. Only OpenAI's GPT-6.1-Sol scored higher, at 76.32. Darwin-180B-RSI-R3 outperformed Google's Gemini-4-Argon and DeepSeek-V4.1-Flash, both of which scored 71.05, as well as xAI's Grok-4.7 (68.42). Among open-weight models — those with publicly available weights — Darwin-180B-RSI-R3 ranked first. It was also the only Korean model to appear in the SuperCLUE rankings.
The model's base is Alibaba's Qwen3.8-Flash-Next. Its parent model, Qwen3.8-Flash, scored 57.89 on the same metric — meaning VIDRAFT's self-improvement process yielded a substantial gain over the starting point.
The architecture is a Mixture-of-Experts (MoE) design with 180 billion total parameters, of which approximately 30 billion are activated during inference. VIDRAFT also released a 4-bit GGUF quantized version (111 GB) that can run on consumer hardware without a dedicated GPU.
The self-improvement method works by having the model solve verifiable problems autonomously. Only solutions the model itself can confirm as correct are recycled back into training data. No human-written answers or externally generated reasoning traces are used in this loop.
However, the model's strengths are concentrated. On SuperCLUE's five-domain composite index, Darwin-180B-RSI-R3 scored 43.08, placing ninth out of 19 models overall and fifth among open-weight models. 라보메모 noted that the model falls short of competitors in areas like long-context processing and instruction following.
The release of Darwin-180B-RSI-R3 is significant for several reasons. First, it demonstrates that a startup — not one of the major AI labs — can produce an open-weight model competitive with closed proprietary systems on specialized reasoning tasks. The fact that VIDRAFT's model surpasses well-resourced competitors from Google and DeepSeek on math and science reasoning, while making its weights publicly available, is a meaningful data point in the ongoing conversation about open vs. closed AI development.
Second, the recursive self-improvement methodology is notable. By training exclusively on the model's own verified outputs rather than human-labeled data or distilled reasoning chains, VIDRAFT is pursuing a path toward increasingly autonomous model improvement — a direction that aligns with broader industry interest in scalable, low-human-intervention training pipelines.
Third, the GGUF quantized version lowers the barrier to access considerably. Researchers, developers, and hobbyists without access to large GPU clusters can run the model locally, which expands the potential community of users and contributors.
The release also arrives amid a broader trend of high-performing open-weight models entering the market. 라보메모's roundup mentioned other examples in the same vein, suggesting that competitive open-weight models are becoming more common across the industry.
Q: What is Darwin-180B-RSI-R3 and who made it?
A: Darwin-180B-RSI-R3 is a large language model developed by VIDRAFT, a Korean AI startup. It was trained using recursive self-improvement (RSI) and released as an open-weight model on November 6, 2026.
Q: How did Darwin-180B-RSI-R3 perform on SuperCLUE's math and science reasoning benchmark?
A: According to VIDRAFT's published figures, the model scored 73.68 on SuperCLUE's November 2026 math and science reasoning sub-index, placing first among all open-weight models and second overall out of 19 evaluated models.
Q: Can Darwin-180B-RSI-R3 run on consumer hardware?
A: Yes. VIDRAFT released a 4-bit GGUF quantized version of the model at 111 GB, which is designed to run on a standard PC without a dedicated GPU.
Source: 라보메모 ラボメモ (일본) (2026-11-07) — original article