Read the headline, not just the score
Benchmark news is usually read through the score. This one should be read through the headline.
At 19:53 on 4 November 2026, SuperCLUE published its November monthly evaluation. The author is the SuperCLUE team and the channel is the official CLUE Chinese Language Understanding Evaluation account. Its headline reads:
"November 2026 Chinese large model benchmark results released! Kimi-K3.1-Pro #1 in China, Korean model Darwin on the board for the first time!"
That is a translation of the Chinese original, which closes with the phrase for "on the board for the first time".
There were two slots in that headline. One went to China's top-scoring model. The other came to us. Out of 19 entrants, the party that chose which model to name in the headline was the party that did the scoring. This was not a press release we sent, and not our summary of someone else's table. The evaluator put our model's name in its own headline.
The scoring design the announcement disclosed
The announcement published this round's weights as numbers, which shows exactly which capability it valued.
| Task | Weight |
|---|---|
| Agentic coding | 30% |
| Math and science reasoning | 24% |
| Hallucination control | 18% |
| Precise instruction following | 18% |
| Ultra-long context | 10% |
This round put the most weight on agentic coding, with math and science reasoning second. In the announcement's own words, that axis "emphasises logical derivation and quantitative analysis". All 19 models sat the same private exam under the same weights.
What happened on the hardest axis
The announcement reports the math and science reasoning result like this:
"GPT-6.1-Sol leads the math and science reasoning board with 76.32, Darwin-180B-RSI-R3 (73.68) takes second, and DeepSeek-V4.1-Flash-1017 and Gemini-4-Argon tie for third at 71.05"
Second of 19. The gap to first is 2.64 points. Below us on that axis sit Gemini-4-Argon, DeepSeek-V4.1-Flash, Grok-4.7 and Qwen3.8-Max-0902. Among open-weight models it is first.
Darwin-180B-RSI-R3 is the only Korean model on the board.
The footnote written only for Darwin
One more line in the announcement applies to our model alone. It appears in the cost-efficiency section.
"*Darwin-180B-RSI-R3 is an open-source deployed model, so API pricing and cost are not tallied.
The open-source address is available at https://huggingface.co/FINAL-Bench/Darwin-180B-RSI-R3"
Two things follow from that footnote. First, the other leading models on the board have their cost efficiency computed from API pricing, and ours is not in that calculation at all. A model you pay per call to use and a model whose weights you download and run on your own hardware sit on entirely different cost axes. Second, the evaluator linked our repository directly in the body of its own announcement, so a reader who sees the score can fetch the weights immediately.
It shares the board with its parent model
Darwin-180B-RSI is a 180 billion parameter mixture-of-experts model with roughly 3 billion 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 and the shared experts through recursive self-improvement: the model writes problems, solves them, verifies the answers and updates its own weights from the verified traces.
Both models are on this board, so the comparison is direct.
| Qwen3.8-Flash | Darwin-180B-RSI-R3 | |
|---|---|---|
| Math and science 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.
Why this document is different
Secondary coverage of the same result already exists. NetEase ran it on 7 November as "Second at 73.68, the open-weight model that left Gemini and Grok behind", and in Korea, ChosunBiz, Economist, Digital Daily and AI Times all reported it on 6 November within hours of each other.
This announcement is the source those pieces drew from. It is the primary document, written by the party that produced the numbers, and the choice of headline was theirs too. Us saying "first among open-weight models" and the evaluator writing "Korean model Darwin on the board for the first time" are the same fact carrying different weight.
Where to check it and what you can run
The announcement is published here, and the full board is at superclueai.com. The weights are at Darwin-180B-RSI-R3 under the Qwen community licence with commercial use permitted, and the 4-bit on-device build POCKET-Darwin-180B-GGUF runs from SSD on a 32GB 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.
What does "on the board for the first time" mean?
It means a Korean model appeared on the SuperCLUE board for the first time. The headline uses the Chinese phrase meaning "appears on the board for the first time".
How do the two models in the headline relate?
Kimi-K3.1-Pro is first among Chinese models with an overall index of 61.46, and Darwin-180B-RSI-R3 is second of 19 on math and science reasoning at 73.68, first among open-weight models. The announcement put both facts in one headline.
Is Darwin first overall?
It is first among open-weight models and second of 19 on math and science reasoning. On the displayed overall index it is 9th of 19 and 5th among open-weight models.
What licence covers the weights?
Qwen Community Licence 1.0, inherited from the model whose expert layers it reuses. Commercial use is permitted.
Top comments (0)