VIDRAFT's 180B Darwin beats 2.8T Kimi in SuperCLUE math and science test

An open-weight model from Korean developer VIDRAFT has taken first place among open-weight models in SuperCLUE's November 2026 evaluation of mathematical and scientific reasoning, finishing ahead of a competitor more than fifteen times its size. Darwin-180B-RSI scored 73.68, second only to OpenAI's GPT-6.1-Sol at 76.32, according to two posts published on dev.to, one of them from VIDRAFT's own lab account.

The November board

Moonshot AI's Kimi-K3.1-Pro, at 2.8 trillion parameters, scored 63.16 — more than ten points behind Darwin, which has 180 billion parameters, roughly one sixteenth of that size. SuperCLUE is one of four evaluators commonly cited for independent testing, alongside LMArena, Artificial Analysis and Stanford HELM. It does not accept developer-submitted scores: the evaluator calls each model directly and grades it on a private test set that changes every month.

The November evaluation spans five tracks — agentic coding, math and science reasoning, hallucination control, precise instruction following, and long context. Darwin's lead is specific to math and science reasoning. On the combined index it ranks fifth among open-weight models and ninth overall out of 19 models.

Merging strengths, not whole models

According to the posts, Darwin names both a model and a creation method. VIDRAFT selects two or more public models as parents, analyses the feed-forward regions where each parent stores its knowledge, extracts the strengths specific to each one, and synthesises them into a single child model. The approach differs from evolutionary model merging, which searches over merge combinations and blends parents wholesale; Darwin first identifies which region holds which strength and transplants selectively. VIDRAFT says the resulting model scores more than 15 points above the comparable public base model in math and science reasoning.

Model-level recursive self-improvement

The synthesised child then goes through model-level RSI. No human-written solutions or reasoning traces are involved: the model writes its own solutions, an automatic verifier checks each answer, and training runs only on the solutions that pass. VIDRAFT describes the weights as improving round after round with no person in the loop.

What one sixteenth the size buys

The size gap is an economics story. A smaller model serves the same workload on fewer machines, which lowers cost per token. VIDRAFT reports server GPU inference at 18,057 tokens per second, concurrent throughput at 2.6 times a standard 4-bit build on the same hardware, and 417 tokens per second on NPUs such as the Korean-made Rebellions ATOM-Max. A POCKET edition runs on a single laptop, keeping inference on-device — a fit, the Korean post argues, for security-sensitive enterprises and public institutions, and a foundation for sovereign AI that reduces reliance on foreign GPUs and cloud providers.

Downloads and claimed records

VIDRAFT's open models have around 5 million cumulative downloads on Hugging Face and more than 1.2 million per month, the posts state. The POCKET family recorded 1,194,702 downloads in the 30 days to 13 November 2026, with POCKET-Darwin-180B-GGUF accounting for 525,554. The Korean post adds that the cumulative total is 6.6 times Sakana AI's figure, and that Alibaba's ModelScope registers VIDRAFT as the only officially certified Korean AI organisation. On Hugging Face benchmark leaderboards, the Darwin family holds first place on 12 boards, including GPQA at 94.44 and perfect scores on AIME 2026 and HMMT.

A caveat on provenance: one of the two posts was authored by VIDRAFT itself. The SuperCLUE score is third-party, but the serving-speed, download and leaderboard figures are vendor-reported and have not been independently verified here.

Why it matters

If a 180B open-weight model can beat a 2.8 trillion-parameter model by ten points under third-party evaluation, capability per parameter is improving faster than raw scale — and that shifts the frontier of what organisations can run on hardware they own. The two techniques behind the result, selective strength merging and verifier-filtered self-training, are in principle reproducible by other open developers, and both reduce dependence on scarce human-labelled reasoning data. If the serving numbers hold up outside the vendor's own tests, they translate directly into lower cost per token and viable on-device deployment.