Korean Press on Darwin's SuperCLUE Result: Chosun Biz, Economist, Digital Daily and AI Times Report #2 in Math and Science Reasoning
aillmbenchmarknews
TL;DR
On November 6, 2026, four Korean news outlets reported that Darwin-180B-RSI-R3, VIDRAFT's open-weight reasoning model, scored 73.68 in mathematical and scientific reasoning on SuperCLUE's November 2026 general leaderboard. That is second among the 19 models evaluated, behind only OpenAI's GPT-6.1-Sol (76.32), and first among open-weight models. Chosun Biz led with the models Darwin passed: Google Gemini, xAI Grok and DeepSeek. Economist Korea and Digital Daily led with the global rank. AI Times added the detail that the model was built on Alibaba's Qwen3.8-Flash-Next and that it also placed 9th on the overall index.
What the four outlets wrote
| Outlet | Headline (translated) | Angle |
|---|---|---|
| Chosun Biz | VIDRAFT's AI model Darwin beats Google, Grok and DeepSeek in math and science reasoning on China's SuperCLUE | Who was beaten |
| Economist Korea | VIDRAFT's Darwin ranks second globally in SuperCLUE math and science reasoning | Global rank, gain over the base model |
| Digital Daily | VIDRAFT's Darwin takes second on a global benchmark, built for complex reasoning | The self-improvement method |
| AI Times | VIDRAFT's Darwin at world number two level in math and science reasoning, ahead of Google and DeepSeek | Method, parent model, overall rank |
All four reported the same core facts from the SuperCLUE board: 19 models, five scoring dimensions, 73.68 in math and science reasoning, second overall and first among open-weight models. Chosun Biz and Economist Korea also noted the earlier result on the Hugging Face GPQA Diamond leaderboard, where the same model holds first place overall at 94.44. AI Times, the most technical of the four, explained recursive self-improvement in its own words: the model writes problems, solves them, checks the answers and updates itself from the verified traces.
The numbers behind the headlines
SuperCLUE is the large-model successor to CLUE, the Chinese language understanding benchmark that began in 2019. It builds a private test set every month, scores five dimensions (agentic coding, mathematical and scientific reasoning, hallucination control, precise instruction following, long context) and publishes the board for Chinese and non-Chinese models alike. Non-Chinese models are shown for reference without an official rank.
Mathematical and scientific reasoning, November 2026:
| Model | Developer | Open weights | Score |
|---|---|---|---|
| GPT-6.1-Sol | OpenAI | No | 76.32 |
| Darwin-180B-RSI-R3 | VIDRAFT | Yes | 73.68 |
| Gemini-4-Argon | No | 71.05 | |
| DeepSeek-V4.1-Flash | DeepSeek | Yes | 71.05 |
| Grok-4.7 | xAI | No | 68.42 |
| Qwen3.8-Max-0902 | Alibaba | No | 68.42 |
| GLM-5.3 | Zhipu AI | Yes | 65.79 |
| Kimi-K3.1-Pro | Moonshot | Yes | 63.16 |
| Qwen3.8-Flash | Alibaba | Yes | 57.89 |
On the overall intelligence index Darwin scored 43.08, 9th of 19 and 5th among open-weight models. The full overall ranking and the second chart are in our earlier report.
Why the parent-model comparison matters
Economist Korea's summary line was that the company reports a large gain in math and science reasoning over the base model. The base model is on the same board. Darwin-180B-RSI inherited its 512 routed experts per layer, the feed-forward networks that hold most of the knowledge, from Alibaba's Qwen3.8-Flash-Next and kept them unchanged. VIDRAFT retrained only the attention layers, the linear attention layers and the shared experts through recursive self-improvement.
| Qwen3.8-Flash (parent) | Darwin-180B-RSI-R3 (child) | |
|---|---|---|
| Math and science reasoning | 57.89 | 73.68 |
| Overall index | 39.81 (15th) | 43.08 (9th) |
Same experts, same knowledge, different reasoning layers: a 15.8 point gain on the hardest dimension and six places on the overall index. That is the result the Korean press picked up as "built for complex reasoning".
What this means for users
Darwin-180B-RSI-R3 is a 180-billion-parameter mixture-of-experts model with about 3 billion parameters active per token. It is released as open weights under the Qwen community licence, which permits commercial use. A 4-bit on-device build, POCKET-Darwin-180B-GGUF, runs from SSD on a laptop with 32 GB of RAM and no datacenter GPU. The scores above were produced by SuperCLUE on its private November test set, not by VIDRAFT.
FAQ
Is this the same result reported on November 5? Yes. SuperCLUE published the board on November 5. The four Korean articles on November 6 are press coverage of that board.
Did Darwin rank first overall? No. It ranked first among open-weight models and second overall in mathematical and scientific reasoning. On the overall five-dimension index it ranked 9th of 19.
Which outlets ran the story? Chosun Biz, Economist Korea, Digital Daily and AI Times, all on November 6, 2026. Links are in the table above.
Where is the primary source? The SuperCLUE board at superclueai.com/homepage. The chart above is a screenshot of that page taken on November 5, 2026.
What is the licence? Qwen community licence 1.0, inherited from the parent model. Commercial use is permitted.