VIDRAFT's open-weight Darwin-180B-RSI-R3 ranks second in SuperCLUE math and science reasoning

VIDRAFT, a Korean AI developer, has released Darwin-180B-RSI-R3, an open-weight large language model that sharpens its own reasoning without fresh human-labeled data. According to a dev.to post describing the release, which credits Korean outlet AI타임스 with the original reporting, the model placed second in math and science reasoning in SuperCLUE's November 2026 evaluation, ahead of entries from Google, DeepSeek, and xAI. Only OpenAI's GPT-6.1 Sol scored higher.

A sparse 180B model built on Qwen

The model is a mixture-of-experts system with 180 billion total parameters, of which roughly 3 billion are active per forward pass. Its starting point is Alibaba's Qwen3.8-Flash-Next, and the final weights ship under an open, commercially usable license. Per the dev.to post, Darwin was the only Korean model among the 19 systems in this SuperCLUE round.

How the self-improvement loop works

Darwin's central claim is its training method, recursive self-improvement (RSI). As described in the post, each round runs a four-step cycle: the model invents new math and science problems, attempts to solve them, critiques its own solutions, and then updates only the weights tied to reasoning. The expert layers that store general knowledge, inherited from the Qwen base, stay frozen throughout, a design meant to keep broad world knowledge intact while reasoning gains compound. The R3 suffix marks the third such round.

The post contrasts this with RLHF and DPO-style fine-tuning, which depend on human-annotated preference or answer data at every improvement cycle. RSI needs no labelers in the loop, which reportedly makes repeated improvement rounds cheaper and more scalable.

The numbers

Citing SuperCLUE's November 2026 math and science results, the post lists GPT-6.1 Sol first at 76.32, Darwin second at 73.68, Gemini 4 Argon and DeepSeek V4.1 Flash tied at 71.05, Grok 4.7 at 68.42, and the Qwen3.8-Flash base at 57.89. That makes Darwin the top open-weight model in the category.

Across all five SuperCLUE domains combined, Darwin scored 43.08, ninth overall and fifth among open-weight models, a 3.27-point gain over the base model's 39.81. The post also reports a 94.44 on the GPQA Diamond evaluation of PhD-level science questions, dated November 3, 2026, an undisclosed improvement on MMLU-Pro, and an 11% shortening of reasoning chains relative to the base model, meaning correct answers arrive with less generated intermediate text.

Caveats worth noting

SuperCLUE primarily benchmarks Chinese-language models, and Darwin's ninth-place overall finish reflects weaker long-context and instruction-following results relative to its reasoning strength, a gap the post says VIDRAFT acknowledges. The practical takeaway is that the model is competitive for math, science, and logic workloads, while general-assistant uses should be tested against a specific workload first. It is also worth flagging that these figures come from a single community post on dev.to, so independent verification of the benchmark claims is still pending.

Availability

The weights are hosted on Hugging Face under VIDRAFT's organization, with cumulative downloads across the Darwin family above 4.68 million as of the report. A compressed variant, POCKET-Darwin-180B, is quantized and distilled to run on consumer CPUs, targeting laptops and desktops without a GPU. The full model still requires GPUs, though sparse activation keeps per-token compute well below what a dense 180B model would demand.

Why it matters

An open-weight model ranking second in a reasoning evaluation dominated by frontier labs narrows the gap between open and closed systems in a domain where proprietary models have led. The training method is the more interesting signal: if self-generated, self-verified problems can substitute for human labels, improvement cycles become cheaper and can repeat without annotation bottlenecks. The trade-off is equally instructive. Targeted self-improvement lifted math and science sharply but left the model mid-pack overall, suggesting that autonomous improvement currently sharpens specific skills rather than raising general capability across the board.