Darwin-180B-RSI Ranks #1 Open-Weight Model in Math and Science Reasoning on SuperCLUE (Nov 2026)

TL;DR

SuperCLUE, China's most widely cited LLM evaluation, published its November 2026 general leaderboard on November 5. Nineteen models were scored, including Gemini-4-Argon, GPT-6.1-Sol, Grok-4.7, Kimi-K3.1-Pro, DeepSeek-V4.1-Flash, GLM-5.3 and Qwen3.8-Max. Darwin-180B-RSI-R3, an open-weight model from VIDRAFT, scored 73.68 in mathematical and scientific reasoning. That is second overall, behind only GPT-6.1-Sol at 76.32, and first among open-weight models. On the overall intelligence index Darwin placed 9th of 19 and 5th among open-weight models. It is the only Korean model on the board.

The result that matters most for how we build models is this: Darwin inherited its expert layers from Qwen3.8-Flash-Next, and the parent model, listed on the same board as Qwen3.8-Flash, placed 15th overall with 57.89 in math and science reasoning. The child outscored the parent by 15.8 points on the same test.

What does SuperCLUE measure?

SuperCLUE is the large-model continuation of CLUE, the Chinese language understanding benchmark that started in 2019. The team published its methodology with researchers from Westlake University (arXiv 2307.15020). Each month it builds a fresh, private test set, so a model cannot gain points by memorising released questions. The general leaderboard scores five dimensions: agentic coding, mathematical and scientific reasoning, hallucination control, precise instruction following, and long context. The overall intelligence index is the average across the five.

Because the questions are in Chinese and the evaluators are independent, SuperCLUE is the scorecard Chinese labs quote when they ship a new model, and it is the one place where US, Chinese and other models are measured on the same Chinese-language paper every month.

Where did Darwin land?

Mathematical and scientific reasoning, November 2026:

Overall intelligence index, same month: Gemini-4-Argon 73.62, GPT-6.1-Sol 71.76, Kimi-K3.1-Pro 61.46, DeepSeek-V4.1-Flash 58.95, GLM-5.3 50.66, Grok-4.7 49.66, GLM-5.3-Flash 47.18, Qwen3.8-Max 43.22, Darwin-180B-RSI-R3 43.08, Hy4-Preview 42.96, Doubao-Seed-2.1-Pro 42.06, MiniMax M3.1-Flash 41.35, Step-5-Preview 41.23, MiniMax-M3 41.17, Qwen3.8-Flash 39.81, LongCat-2.5-Preview 39.59, MiMo-V2.6-Flash 36.40, Spark-X2.6 35.39, MiMo-V2.6-Pro 34.57.

Darwin is 9th overall and 5th among open-weight models, behind Kimi-K3.1-Pro, DeepSeek-V4.1-Flash, GLM-5.3 and GLM-5.3-Flash. SuperCLUE lists non-Chinese models for reference without an official rank, and Darwin is not in the price chart because VIDRAFT did not submit API pricing.

How was Darwin-180B-RSI built?

Darwin-180B-RSI is a 180-billion-parameter mixture-of-experts model. Each layer holds 512 routed experts, and about 3 billion parameters are active per token, so the model has the knowledge of a very large network and the serving cost of a mid-sized one.

The expert layers, the feed-forward networks that store most of the factual knowledge, were inherited from Alibaba's Qwen3.8-Flash-Next and kept unchanged. What VIDRAFT rebuilt is the part of the network that does the reasoning: the full attention layers, the linear attention layers and the shared experts. Those weights were trained through recursive self-improvement (RSI). The model writes its own problems, solves them, verifies the answers, and the verified traces are used to update the model. No human-written answer set is involved, so the loop can run round after round and keep improving.

That is why the comparison with the parent is informative. Same expert layers, same knowledge, different reasoning layers. On SuperCLUE the parent, Qwen3.8-Flash, scored 57.89 in math and science reasoning and placed 15th overall. Darwin scored 73.68 and placed 9th.

Two days earlier, on November 3, the same model took first place overall on the Hugging Face GPQA Diamond leaderboard with 94.44. GPQA Diamond is a graduate-level physics, chemistry and biology test in English. Scoring at the top of an English science test and a Chinese math and science test in the same week is evidence that the gain is in reasoning itself, not in a test-specific fine-tune.

What should you take from this?

Three things.

First, open weights are not a ceiling. On this board the open-weight model with the best math and science score beat every closed model except one.

Second, you do not need to pretrain from scratch to move a model's reasoning. Inheriting expert layers and retraining the attention and shared-expert layers through a self-improvement loop moved a 180B model from 15th to 9th overall and from 57.89 to 73.68 on the hardest dimension.

Third, the model is available. Darwin-180B-RSI-R3 is released as open weights under the Qwen community licence, which permits commercial use. A compressed on-device build, POCKET-Darwin-180B-GGUF, runs on a laptop without a GPU.

FAQ

Is SuperCLUE an official or government benchmark?
No. It is an independent benchmark run by the CLUE team, which has published Chinese NLP benchmarks since 2019. It is widely cited by Chinese labs and media because the test set is private and refreshed monthly.

Why is Darwin's overall rank only 9th if it is 2nd in reasoning?
The overall index averages five dimensions. Darwin's long-context score (30.32) and precise instruction following (20.69) are far below its reasoning score, and those pull the average down.

Did VIDRAFT submit Darwin to SuperCLUE?
SuperCLUE selects public models for its monthly board. VIDRAFT served the model through an API endpoint during the evaluation window.

What is the licence?
Qwen community licence 1.0, inherited from the parent model. Commercial use is permitted; the licence text ships with the weights.

Where can I verify the numbers?
The leaderboard is at superclueai.com/homepage. The two charts above are screenshots of that page from November 5, 2026.

Further reading

• VKORE-325M: pretraining a language model on an inference-only NPU
• ZTC: a decision model that answers in one forward pass