How a 180B Open Model Outscored a 2.8T Model in Math and Science Reasoning: Strength Merging Plus Model-Level RSI

TL;DR

Darwin-180B-RSI, an open-weight model from Korea's VIDRAFT, scored 73.68 in mathematical and scientific reasoning on SuperCLUE's November 2026 evaluation. That is first among open-weight models and second overall, directly behind OpenAI's GPT-6.1-Sol (76.32). The largest model in the same evaluation, Moonshot AI's Kimi-K3.1-Pro at 2.8 trillion parameters, scored 63.16. Darwin has 180 billion parameters, about one sixteenth of that size, and finished more than 10 points ahead.

This post explains the two techniques behind the result and what the size difference means for serving cost.

The full ranking

Why SuperCLUE counts as third-party verification

SuperCLUE does not accept scores submitted by model developers. The evaluator calls each model directly and grades it on a private test set that changes every month. SuperCLUE, LMArena, Artificial Analysis and Stanford HELM are the four evaluators commonly cited for this style of independent testing.

The November evaluation covers five tracks: agentic coding, math and science reasoning, hallucination control, precise instruction following, and long context. On the combined index Darwin ranks fifth among open-weight models and ninth overall out of 19 models.

Technique 1: merge the strengths, not the whole model

Darwin is also the name of VIDRAFT's model creation method. It selects two or more public models as parents, analyzes the feed-forward (FFN) regions where each parent stores its knowledge, extracts the strengths specific to each one, and synthesizes them into a single child model.

Evolutionary model merging searches over merge combinations and blends the parents. Darwin first identifies which region holds which strength and then transplants selectively. The result is a new model built from chosen strengths, and in math and science reasoning it scores more than 15 points above the comparable public base model.

Technique 2: model-level recursive self-improvement

The synthesized child then goes through model-level RSI. No human-written solutions or reasoning traces are used. The model writes its own solutions, an automatic verifier checks each answer, and the model trains on the solutions that pass. The weights themselves improve their reasoning ability, round after round, with no person in the loop.

What one sixteenth the size buys you

Metric	Value
Server GPU inference speed	18,057 tokens per second
Concurrent request throughput	2.6x a standard 4-bit build on the same hardware
NPU inference speed	417 tokens per second

A smaller model serves the same workload on fewer machines, which lowers the cost per token. Darwin also runs on Korean-made NPUs such as Rebellions ATOM-Max through VIDRAFT's VKORE core technology, and the POCKET edition runs on a single laptop as on-device AI.

Public records

On the official Hugging Face benchmark leaderboards, the Darwin family currently holds first place on 12 boards. These include GPQA at 94.44, and perfect scores on AIME 2026 and HMMT.

Adoption follows the same curve. VIDRAFT's open models have about 5 million cumulative downloads on Hugging Face and more than 1.2 million per month. The POCKET family alone recorded 1,194,702 downloads in the 30 days to November 13, 2026, with POCKET-Darwin-180B-GGUF at 525,554.

Safety and the rest of the stack

ViKernel is VIDRAFT's weight-native operating system: rules inside the model weights judge every action the AI is about to take, block dangerous ones, and keep a kill switch available to a human at all times. AETHER is a model trained from scratch and released under Apache 2.0.