Open-weight Darwin-180B MoE reportedly takes second on SuperCLUE math and science test

Open-weight model lands near the top of a Chinese reasoning benchmark

According to a dev.to report citing Chinese outlet NetEase, Darwin-180B-RSI-R3, a Mixture-of-Experts language model from South Korean startup VIDRAFT, scored 73.68 on SuperCLUE's November 2026 math and science reasoning dimension. That placed it second out of 19 evaluated models and first among open-weight entries, with only GPT-6.1-Sol, at 76.32, ahead of it. The same table put it above Gemini, Grok and DeepSeek on that specific dimension, the report says.

What the model is

Darwin-180B-RSI-R3 has 180 billion parameters in total but activates only about 3 billion per token at inference. It uses Alibaba's Qwen3.8-Flash-Next as its base, inheriting the 512 experts per layer whose feed-forward blocks hold most of the model's stored world knowledge. VIDRAFT did not retrain those knowledge-bearing blocks. Instead, per the report, it retrained only the attention layers, the linear-attention layers and the shared expert layers, the components most tied to multi-step reasoning. The weights are published under the Qwen community license, which the report says permits commercial use, and can be found on VIDRAFT's organisation page on Hugging Face.

A self-improvement training loop

Korean media coverage cited in the dev.to post describes the retraining as a loop: the model writes its own problems, drafts candidate answers, checks them internally, and uses that feedback signal to update its weights. The approach resembles self-play or self-distillation techniques familiar from reasoning research, applied narrowly to the layers that govern how information is routed rather than what the model knows. The reported payoff is a 15.8-point gain over the Qwen base model on the reasoning dimension, plus a six-place climb on SuperCLUE's composite ranking.

The numbers, with caveats

SuperCLUE grew out of the CLUE Chinese language evaluation project launched in 2019. It issues a fresh, private test set every month and scores models across five dimensions: agentic coding, math and science reasoning, hallucination suppression, instruction following, and long-context understanding. The November set was administered on 5 November 2026, and the report stresses the scores were produced by SuperCLUE itself rather than self-reported by VIDRAFT, on test items that were not public in advance.

Two caveats stand out. First, Darwin is a narrow specialist: on SuperCLUE's five-dimension composite index it scored 43.08, ninth of 19 models overall and fifth among open-weight ones. The report also cites a GPQA Diamond score of 94.44, ranked first on the Hugging Face leaderboard at the time of reporting. Second, SuperCLUE lists non-Chinese models such as Gemini and Grok on the same table for reference but excludes them from its official ranking, so the second-place headline should be read with that framing in mind. The story also currently rests on a single report relayed through dev.to, so the usual caution for one-source claims applies.

A laptop-sized variant

The full-precision 180B model is not consumer hardware territory, but VIDRAFT also released POCKET-Darwin-180B-GGUF, a 4-bit quantized GGUF build intended to run from an SSD on a laptop with 32 GB of RAM and no GPU. The dev.to post notes that the release coverage did not document a specific API endpoint, repository URL or installation command, so anyone intending to deploy it should confirm access details through VIDRAFT's official channels.

Why it matters

Two things make this more than another leaderboard blip, assuming the numbers hold up. One is efficiency: a model that activates roughly 3 billion parameters per token, small enough to run quantized on an ordinary laptop, landed within about 2.6 points of a frontier closed model on a private reasoning test it could not have seen in advance. That fits the broader pattern of open-weight models closing the gap with proprietary ones at a fraction of the serving cost. The other is the training method: rather than retraining a giant model end to end, VIDRAFT reportedly froze the layers that store knowledge and retrained only the reasoning-related ones using supervision the model generated itself. If that layer-selective approach generalises, it points to a comparatively cheap recipe for specialising large models. For now, though, one benchmark dimension from one independently scored test is an interesting data point rather than proof.