A new open-weight AI model named Darwin-180B-RSI, developed by Korea's VIDRAFT, has achieved a score of 73.68 in mathematical and scientific reasoning on the SuperCLUE evaluation. This performance places it second overall, behind OpenAI's GPT-6.1-Sol, and significantly outperforms larger models like Moonshot AI's Kimi-K3.1-Pro, which scored 63.16 despite having 2.8 trillion parameters compared to Darwin's 180 billion. The model's success is attributed to two key techniques: strength merging, which synthesizes knowledge from multiple parent models, and recursive self-improvement (RSI) for reasoning ability without human intervention. These advancements allow Darwin to offer competitive performance at a fraction of the size and cost, with versions optimized for various hardware, including laptops. AI
IMPACT Demonstrates that smaller, efficiently trained models can rival much larger ones in complex reasoning tasks, potentially lowering AI deployment costs.
RANK_REASON The item details a new model release and its performance on benchmarks, including novel techniques used in its development. [lever_c_demoted from research: ic=1 ai=1.0]
- AIME 2026
- ATOM-Max
- Darwin-180B-RSI
- GPQA: A Graduate-Level Google-Proof Q&A Benchmark
- HMMT
- Hugging Face
- Moonshot AI
- OpenAI
- Rebellions
- SuperCLUE
- VIDRAFT
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
