A Korean AI model lands on China's SuperCLUE leaderboard for the first time

What SuperCLUE announced

SuperCLUE, a Chinese benchmark that scores AI models on a private test set, released its November 2026 monthly evaluation on 4 November. According to a post on dev.to written from the perspective of Darwin's developer, 19 models sat the same exam under the same weights.

The announcement's headline named two things: Kimi-K3.1-Pro as the top model in China, and a Korean model, Darwin, appearing on the board for the first time. The post's author stresses that the evaluator chose that wording, not the model's makers.

That is worth noting with care. The dev.to post comes from the team behind Darwin, so it is an interested party describing a result in its favor. The scores it quotes are attributed to SuperCLUE's announcement, and the full board is published on SuperCLUE's site for anyone to check.

How the exam was weighted

SuperCLUE published the weights for this round. Agentic coding, where a model writes and runs code to finish a task on its own, counted for 30%. Math and science reasoning counted for 24%.

The remaining weight was split between hallucination control (18%), which measures how often a model makes things up, precise instruction following (18%), and ultra-long context (10%), which tests handling of very long documents.

Because reasoning was the second-heaviest category, a strong result there matters for the overall score, though it was not enough to put Darwin near the top overall.

Where Darwin placed

On math and science reasoning, GPT-6.1-Sol led with 76.32. Darwin-180B-RSI-R3 took second with 73.68, a gap of 2.64 points. DeepSeek-V4.1-Flash-1017 and Gemini-4-Argon tied for third at 71.05.

Grok-4.7 and Qwen3.8-Max-0902 also sit below Darwin on that axis. Among open-weight models, meaning those whose files can be downloaded and run by anyone, it is first on this measure.

The overall picture is more modest. On the displayed overall index Darwin is 9th of 19 and 5th among open-weight models. It is also the only Korean model on the board.

How the model was built

Darwin-180B-RSI is a mixture-of-experts model, a design where many specialist sub-networks exist but only a few activate for each piece of text. It has 180 billion parameters, with roughly 3 billion active per token.

According to the post, it copied its 512 routed experts per layer from Alibaba's Qwen3.8-Flash-Next and rebuilt other parts, including the attention layers, through recursive self-improvement. In that process the model writes problems, solves them, checks the answers, and updates itself from the verified work.

Both Darwin and Qwen3.8-Flash appear on the board, which allows a direct comparison. On math and science reasoning Qwen3.8-Flash scored 57.89 against Darwin's 73.68. On the overall index it scored 39.81 (15th) against 43.08 (9th).

Cost, licence and access

SuperCLUE added a footnote for Darwin: because it is deployed as an open-source model, API pricing and cost were not tallied. Other models on the board have cost efficiency computed from per-call pricing, so the two groups are not compared on the same cost axis.

The announcement also links to the model's Hugging Face page. The weights fall under the Qwen Community Licence, which permits commercial use. A 4-bit version, compressed to run on smaller hardware, is said to run from SSD on a 32GB laptop.

As dev.to reported, Korean outlets including ChosunBiz, Economist, Digital Daily and AI Times covered the result on 6 November, and NetEase followed on 7 November.