Darwin-180B-RSI-R3: Open-Weight MoE Scores #2 on SuperCLUE Math & Science Reasoning, Beating Gemini and Grok

TL;DR: Darwin-180B-RSI-R3 is a 180-billion-parameter Mixture-of-Experts model from VIDRAFT that scored 73.68 on the SuperCLUE November 2026 math and science reasoning dimension — ranking #2 overall and #1 among open-weight models on an independently administered, non-public test. It activates only ~3B parameters per token, ships weights under a Qwen community license permitting commercial use, and has a 4-bit quantized variant that runs on a 32 GB RAM laptop without a GPU.

What it is

Darwin-180B-RSI-R3 is a Mixture-of-Experts (MoE) language model built on top of Alibaba's Qwen3.8-Flash-Next as a base. Key facts from the source:

• Parameter count: 180B total, with roughly 3B activated per token at inference time
• Architecture lineage: Inherits the expert layers from Qwen3.8-Flash-Next — 512 experts per layer, with the bulk of world knowledge stored in those feed-forward network (FFN) blocks
• What changed vs. the base: Only the attention layers, linear-attention layers, and shared expert layers were retrained using a self-improvement process; the FFN expert blocks were kept intact
• License: Qwen community license, open-weight, commercial use permitted
• Edge variant: POCKET-Darwin-180B-GGUF — a 4-bit quantized GGUF version designed to run from SSD on a 32 GB RAM consumer laptop, no GPU required
• Developer: VIDRAFT (South Korean Pre-AGI AI startup)

The model was independently evaluated by SuperCLUE on November 5, 2026. SuperCLUE is a Chinese LLM benchmark that evolved from the 2019 CLUE (Chinese Language Understanding Evaluation) project. It issues fresh, non-public test sets each month and scores models across five dimensions: agentic coding, math & science reasoning, hallucination suppression, precise instruction following, and long-context understanding.

How it works

The core idea behind Darwin-180B-RSI-R3 is surgical retraining: instead of fine-tuning an entire large model, VIDRAFT preserved the FFN expert layers — which encode the model's factual and domain knowledge — and exclusively retrained the components most responsible for multi-step reasoning: the attention layers, linear-attention layers, and shared experts.

The retraining mechanism is described in Korean media coverage as a self-improvement loop: the model generates its own problems, produces candidate answers, verifies those answers internally, and uses that feedback signal to update weights. This is conceptually similar to self-play or self-distillation approaches common in the reasoning research literature, applied specifically to the layers governing how the model attends to and routes information rather than what it knows.

The practical consequence of this design choice: the same 512-expert FFN knowledge base as the parent model, but with reasoning-focused layers optimized through iterative self-generated supervision.

Benchmarks & results

All numbers below come directly from the SuperCLUE November 2026 leaderboard (published 2026-11-05) and are independently administered — not self-reported by VIDRAFT.

The model ranked above Google Gemini, xAI's Grok, and DeepSeek on the math & science reasoning dimension. The only model ahead of it was GPT-6.1-Sol at 76.32.

Important context: SuperCLUE places non-Chinese models on the same table for reference but does not include them in its official ranking. The 73.68 score is from a held-out, non-public November test set — not a contamination-prone static benchmark.

How to try it

The source article confirms that model weights are publicly available. Based on the information reported:

• Weights: Released under the Qwen community license (commercial use allowed). Check the VIDRAFT organization page on Hugging Face for Darwin-180B-RSI-R3.
• Edge/CPU variant: POCKET-Darwin-180B-GGUF (4-bit quantized) is designed for 32 GB RAM laptops running inference from SSD — no GPU required. Look for it on Hugging Face under VIDRAFT's model hub.

The source does not publicly document a specific API endpoint, GitHub repository URL, or installation command for this release. Check VIDRAFT's official channels for the latest access instructions before deployment.

FAQ

Q: Why does the model score #2 on reasoning but only #9 overall on SuperCLUE?
A: Darwin-180B-RSI-R3 was specifically optimized for math and science reasoning. The five-dimension SuperCLUE composite also includes agentic coding, hallucination suppression, instruction following, and long-context tasks — areas where the model was not specifically retrained. The result is a narrow specialist that excels on its target dimension while sitting mid-table overall.

Q: Is retraining only the attention layers a proven approach, or experimental?
A: The approach is conceptually grounded in existing research showing that FFN layers store factual knowledge while attention layers govern reasoning and information routing. The +15.8-point gain on an independently scored, non-public benchmark is a concrete external data point, though a single benchmark result should always be interpreted alongside broader evaluation.

Q: Can I run the full 180B model on consumer hardware?
A: Not the full-precision model. The POCKET-Darwin-180B-GGUF 4-bit quantized variant is specifically designed for 32 GB RAM laptops running from SSD without a GPU, according to the source coverage.

Q: Does the Qwen community license allow me to use this in a commercial product?
A: The source states commercial use is permitted under the Qwen community license. Always review the full license text in the model repository before production deployment.

Originally reported by 넷이즈 网易 (중국) (2026-11-07) — source article.