Open-weight Darwin-180B tops SuperCLUE math and science, ahead of Gemini and DeepSeek

An open-weight model with 180 billion parameters has taken first place among open-weight systems on SuperCLUE's November 2026 math and science leaderboard. Darwin-180B-RSI-R3, built by Korean startup VIDRAFT, scored 73.68 in the category — second overall among 19 models and ahead of Gemini-4-Argon and DeepSeek-V4.1-Flash, which tied at 71.05. The figures come from a dev.to writeup published November 10 that summarizes an earlier report by USB.RU.

Where it placed

The only model ahead of Darwin on the math and science axis was GPT-6.1-Sol at 76.32. Kimi-K3.1-Pro scored 63.16 in the same category, and the base model Darwin was fine-tuned from, Qwen3.8-Flash-Next, managed 57.89 — a gap of nearly 16 points attributable to the fine-tuning process. Darwin was also the only Korean model on the table, according to the writeup.

The general leaderboard tells a more modest story. Darwin's overall SuperCLUE score of 43.08 ranked ninth of 19, ahead of the base model's 39.81 but far from the leaders. Two weaknesses stand out in the leaderboard data: precise instruction-following at 20.69 and long-context tasks at 30.32, both well below the front of the field. On current evidence, Darwin is a math and science specialist rather than a general-purpose assistant.

How the training loop works

VIDRAFT refined the Qwen3.8-Flash-Next base with a method it calls Recursive Self-Improvement (RSI). The loop has four steps: the model generates solutions to a problem set, answers are checked against ground-truth keys, only the reasoning traces that led to correct answers are kept, and the model is fine-tuned on that filtered signal. The cycle then repeats with the updated weights. Unlike supervised fine-tuning on a fixed dataset, the training signal strengthens as the model improves, and no pre-curated chain-of-thought corpus is required.

The updates were deliberately narrow. Only attention paths and shared experts were modified; the base model's 512 experts, its router, and its visual encoder were left untouched. VIDRAFT says this limits catastrophic forgetting and preserves the base model's general capabilities while concentrating gains in sequential reasoning. One measurable side effect: average reasoning chain length on MMLU-Pro fell from 4,320 tokens to 3,833 — a reduction of roughly 11% — meaning the model reaches correct answers with shorter chains of thought after training.

Treat some scores as provisional

The model card reports GPQA Diamond at 94.44, MMLU-Pro at 88.12, and perfect scores on both AIME 2026 and HMMT February 2026. The dev.to writeup notes that USB.RU stated independent third-party verification of the GPQA and similar figures has not yet been performed, so those numbers should be read as self-reported until externally replicated.

Downloading and running it

The weights are public on Hugging Face in the FINAL-Bench/Darwin-180B-RSI repository under the Qwen Community 1.0 license. A 4-bit quantized GGUF variant, POCKET-Darwin-180B-GGUF, occupies 111 GB. The cited minimum specification is 8 GB of GPU VRAM alongside 32 GB of system RAM, and CPU-only inference is reported at 18–21 tokens per second — putting a 180B-class model within reach of local experimentation without data-center hardware.

Why it matters

Darwin's result is a data point for how far open-weight models have closed the gap on specialized axes: a freely downloadable fine-tune outscoring Gemini and DeepSeek on a math and science benchmark would have been implausible not long ago. The RSI method is also notable as a way to sharpen reasoning without a curated chain-of-thought dataset, and the accessible quantized release widens the pool of researchers who can study it. Two caveats temper the headline: the ninth-place overall score shows the model trades breadth for depth, and several of its strongest numbers remain self-reported pending independent verification.