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

Darwin-180B-RSI-R3 is a 180-billion-parameter Mixture-of-Experts language model developed by VIDRAFT, a South Korean Pre-AGI AI startup. The model scored 73.68 on the SuperCLUE November 2026 math and science reasoning dimension, placing it #2 overall and #1 among open-weight models, according to the SuperCLUE November 2026 leaderboard published on 2026-11-05.

Darwin-180B-RSI-R3 is built on top of Alibaba's Qwen3.8-Flash-Next architecture and inherits the expert layers from that base. However, only the attention layers, linear-attention layers, and shared expert layers were retrained using VIDRAFT's self-improvement process. The feed-forward network expert blocks remained untouched to preserve the model's factual and domain knowledge.

The retraining process, described as a self-improvement loop, involved the model generating its own problems, producing candidate answers, verifying those answers internally, and using that feedback signal to update weights. This targeted retraining allowed the model to optimize the reasoning-focused layers while maintaining the original factual knowledge.

The model uses the Qwen community license, which permits commercial use. VIDRAFT has released the model weights under this license, making it available for download on Hugging Face. Additionally, a 4-bit quantized version called POCKET-Darwin-180B-GGUF is available for running on a 32 GB RAM laptop without a GPU.

When compared to the base Qwen3.8-Flash-Next model, Darwin-180B-RSI-R3 achieved a 15.8-point improvement in math and science reasoning scores, moving up six positions in the SuperCLUE ranking. The only model that outperformed Darwin-180B-RSI-R3 in the math and science reasoning dimension was GPT-6.1-Sol with a score of 76.32.

While Darwin-180B-RSI-R3 excels in reasoning tasks, it ranks #9 overall on the SuperCLUE benchmark, which assesses models across five dimensions, including agentic coding, hallucination suppression, precise instruction following, and long-context understanding. These other dimensions were not the focus of the self-improvement process.

The source article confirms that the full 180B model weights are publicly available, and the POCKET-CPU variant is designed for consumer-level hardware without requiring a GPU. However, specific API endpoints, GitHub repositories, or installation commands for this release are not publicly documented, so users should consult VIDRAFT's official channels for access instructions.