Why does SuperCLUE carry weight?
Because it changes the exam every month and keeps it private.
SuperCLUE is the large-model successor to CLUE, the Chinese language understanding benchmark that started in 2019 and became the reference board for Chinese natural language processing. For large language models it is the most widely cited evaluation in China. Each month it builds a fresh private test set, so a model cannot climb by memorising published questions. It scores five axes: agentic coding, mathematical and scientific reasoning, hallucination control, precise instruction following, and ultra-long context. It places proprietary and open-weight systems in the same table, which is what makes the board readable across the tier divide.
The November 2026 board covers 19 frontier models. Scores below are SuperCLUE's, measured on its private November set.
Who leads each axis
| Axis | Highest displayed score | Leading open-weight model |
|---|---|---|
| Agentic coding | GPT-6.1-Sol 69.23 | DeepSeek-V4.1-Flash-1017 65.38 |
| Mathematical and scientific reasoning | GPT-6.1-Sol 76.32 | Darwin-180B-RSI-R3 73.68 |
| Hallucination control | Gemini-4-Argon 83.19 | Kimi-K3.1-Pro 67.78 |
| Precise instruction following | Gemini-4-Argon and GPT-6.1-Sol 68.97 | Kimi-K3.1-Pro 44.83 |
| Ultra-long context | Gemini-4-Argon 95.63 | Kimi-K3.1-Pro 87.25 |
No single model swept November. Category leadership split between two proprietary systems, and three different open-weight models took the open lead in different axes.
The hardest axis
Mathematical and scientific reasoning is where a board separates pattern completion from working through a problem. The November table:
| 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 |
Darwin-180B-RSI-R3 is first among open-weight models and second of all 19, 2.64 points behind GPT-6.1-Sol and ahead of Gemini-4-Argon, DeepSeek-V4.1-Flash, Grok-4.7 and Qwen3.8-Max. It is the only Korean model on the board.
How the overall index reads
SuperCLUE assigns a numbered rank only to models inside its official scope. International entries appear with a dash in that column as a classification convention, not a performance judgment, so the displayed overall scores and the numbered ranks answer two different questions.
| Model | Overall index | Official numbered rank |
|---|---|---|
| Gemini-4-Argon | 73.62 | dash |
| GPT-6.1-Sol | 71.76 | dash |
| Kimi-K3.1-Pro | 61.46 | No. 1 |
| DeepSeek-V4.1-Flash-1017 | 58.95 | No. 2 |
| GLM-5.3 | 50.66 | No. 3 |
| Darwin-180B-RSI-R3 | 43.08 | dash |
By displayed overall score Darwin places 9th of 19 and 5th among open-weight models.
Same experts, rebuilt reasoning
Darwin-180B-RSI is a 180 billion parameter mixture-of-experts model with about 3 billion parameters active per token. It inherited its 512 routed experts per layer from Alibaba's Qwen3.8-Flash-Next byte for byte and rebuilt only the attention layers, the linear attention layers and the shared experts through recursive self-improvement: the model writes problems, solves them, verifies the answers and updates itself from the verified traces.
Both models sit on this board, so the comparison is direct.
| Qwen3.8-Flash | Darwin-180B-RSI-R3 | |
|---|---|---|
| Mathematical and scientific reasoning | 57.89 | 73.68 |
| Overall index | 39.81 (15th) | 43.08 (9th) |
Same stored knowledge, rebuilt reasoning path: 15.79 points on the hardest axis and six places on the overall index.
Where to check and what to run
The board is published at superclueai.com. The weights are at Darwin-180B-RSI-R3 under the Qwen community licence, commercial use permitted, and the 4-bit on-device build POCKET-Darwin-180B-GGUF runs from SSD on a 32 GB laptop.
FAQ
Who produced these scores?
SuperCLUE, on its private November 2026 test set. VIDRAFT is an evaluated model on that board, not the operator of the benchmark.
Why does a monthly private test set matter?
A public question set leaks into training data over time. Rebuilding the exam each month and keeping it private means a score reflects the model rather than its exposure to the questions.
Is Darwin-180B-RSI-R3 first overall?
It is first among open-weight models and second of 19 in mathematical and scientific reasoning. By displayed overall index it is 9th of 19 and 5th among open-weight models.
Why do Gemini, GPT and Darwin show a dash in the rank column?
SuperCLUE numbers only models inside its official scope. The dash marks that classification, and those models still appear in every category comparison with their scores.
What licence covers Darwin's weights?
Qwen community licence 1.0, inherited from the model whose expert layers it reuses. Commercial use is permitted.
Top comments (0)