A Korean open-weight model just outscored Google Gemini and DeepSeek — here's what happened.
TL;DR: VIDRAFT, a Korean Pre-AGI AI startup, announced on November 6, 2026 that its open-weight reasoning model Darwin-180B-RSI scored 73.68 points in the math and science reasoning category of SuperCLUE's November evaluation, placing second overall among all models assessed. The only model to rank higher was OpenAI's GPT-6.1-Sol at 76.32 points. Darwin-180B also previously topped the Hugging Face GPQA Diamond leaderboard with a score of 94.44.
VIDRAFT's Darwin-180B reasoning model has claimed a remarkable second-place finish on a globally competitive AI benchmark, the Seoul-based Korean AI startup confirmed on November 6, 2026. The result, published as part of SuperCLUE's November comprehensive evaluation, positions Darwin-180B ahead of several flagship models from Google, DeepSeek, and xAI in math and science reasoning — a domain widely regarded as one of the most demanding tests of large language model capability.
VIDRAFT reported that Darwin-180B-RSI — the RSI standing for Recursive Self-Improvement, the training methodology applied — earned 73.68 points in SuperCLUE's math and science reasoning subcategory. That score placed it second among all 19 models evaluated, trailing only OpenAI's GPT-6.1-Sol, which topped the chart at 76.32 points.
The Darwin model outperformed a notable group of competitors in that subcategory: Google Gemini 4-Argon (71.05), DeepSeek V4.1 Flash (71.05), and Grok 4.7 (68.42). Among open-weight models specifically, Darwin-180B surpassed Kimi K3.1 Pro (63.16) and GLM-5.3 (65.79). Darwin was also the only Korean model to participate in the November evaluation.
In the overall composite score across all five assessment categories — which include agent coding, math and science reasoning, hallucination suppression, instruction following, and long-context handling — Darwin-180B scored 43.08 points, ranking ninth out of 19 total models and fifth among open-weight models.
This SuperCLUE result follows another strong showing just days earlier: on November 3, VIDRAFT announced that Darwin-180B had ranked first on the Hugging Face GPQA Diamond leaderboard with a score of 94.44. The GPQA Diamond benchmark uses doctoral-level questions across physics, chemistry, and biology to assess scientific reasoning.
SuperCLUE itself was established in 2023 and is built on the CLUE framework, a Chinese-language LLM evaluation system. It uses undisclosed test items to prevent benchmark contamination and is widely referenced for comparing AI model performance among major developers in China and beyond.
On the technical side, VIDRAFT built Darwin-180B on top of Alibaba's Qwen 3.8-FlashNext as a base model, then applied its proprietary RSI training approach. Rather than relying on externally provided labeled answer data, the method has the model autonomously generate problems, attempt solutions, and iteratively verify its own outputs — a self-supervised loop designed to improve reasoning without human-annotated ground truth. The team preserved the base model's feed-forward expert layers while adjusting attention, linear attention, and shared expert layer weights through this self-improvement process.
The result is a 180-billion-parameter Mixture-of-Experts architecture that activates only around 3 billion parameters during inference, keeping computational overhead low. For context, the base model (Qwen 3.8-FlashNext) scored 57.89 in SuperCLUE's math and science reasoning category and placed 15th overall at 39.81 — meaning Darwin-180B's RSI training produced measurable, concrete gains on both metrics.
VIDRAFT also offers a compressed on-device variant called Pocket Darwin-180B, which is designed to run on personal computers and laptops without dedicated GPU hardware. The full model weights are publicly released, permitting both research and commercial use. Cumulative downloads across Hugging Face have exceeded 4.68 million.
The company's model development work received support from Korea's National IT Industry Promotion Agency (NIPA) under the Ministry of Science and ICT, through a GPU resource provision program. VIDRAFT received an "Excellent" rating during the program's mid-term review in June of this year and has continued releasing and maintaining open models following the conclusion of that project.
VIDRAFT CEO Kim Min-sik commented that, aside from one American model, no other system outscored Darwin in math and science reasoning — and that this result came from a Chinese-language examination administered by a Chinese evaluation body, adding a layer of significance to the outcome. He also highlighted that Darwin's development relied on the model's ability to learn through its own internal process.
Benchmark placements at this level carry real weight in the AI industry: they inform enterprise procurement decisions, signal competitive positioning, and attract developer and research interest. For a Korean startup to rank second globally — ahead of frontier models from Google, DeepSeek, and xAI — in a rigorous, anti-contamination evaluation is a meaningful signal about both the RSI methodology and VIDRAFT's broader research trajectory. The fact that Darwin-180B is fully open-weight and runs on consumer hardware also makes it accessible to a wide global developer community.
Q: What is VIDRAFT's Darwin-180B and what benchmark did it just win?
A: Darwin-180B-RSI is an open-weight large language model developed by Korean AI startup VIDRAFT. In SuperCLUE's November 2026 evaluation, it ranked second globally in math and science reasoning with a score of 73.68, and it also topped the Hugging Face GPQA Diamond leaderboard at 94.44 points.
Q: How does VIDRAFT's Recursive Self-Improvement (RSI) training work?
A: RSI is a self-supervised training approach in which the model generates its own problems, produces candidate solutions, and verifies them iteratively — without relying on externally provided labeled answer data. VIDRAFT applied this method to fine-tune specific layers of the Qwen 3.8-FlashNext base model while preserving its expert feed-forward layers.
Q: Can Darwin-180B be used by developers or run on regular computers?
A: Yes. VIDRAFT releases Darwin-180B as a fully open-weight model, allowing both research and commercial use. The company also offers Pocket Darwin-180B, a compressed variant that runs on standard personal computers and laptops without a dedicated GPU.
Source: 디지털데일리 (2026-11-06) — original article