VIDRAFT's 180 billion parameter model, Darwin-180B-RSI, has achieved a top ranking in SuperCLUE's November 2026 evaluation for mathematical and scientific reasoning. The model outperformed Moonshot AI's Kimi K3.1 Pro, which has 2.8 trillion parameters, and secured second place overall, only behind OpenAI's GPT-6.1 Sol. This achievement is attributed to VIDRAFT's innovative techniques of model synthesis, combining strengths from various parent models, and recursive self-improvement, where the model generates and verifies its own solutions to enhance its weights. AI
IMPACT Demonstrates that advanced training techniques can enable smaller models to outperform larger ones, potentially reducing computational costs for high-performance AI.
RANK_REASON A smaller model outperforming a significantly larger one on key benchmarks, attributed to novel training techniques. [lever_c_demoted from significant: ic=1 ai=1.0]
- Darwin-180B-RSI
- GPT-6.1 Sol
- Hugging Face
- Kimi K3.1 Pro
- Kim Minsik
- Moonshot AI
- OpenAI
- POCKET-Darwin-180B-GGUF
- Seoul Shinmun
- SuperCLUE
- The Korea Economic Daily
- VIDRAFT
- ZDNet Korea
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →