Self-taught Darwin-180B-RSI tops Swiss law-exam benchmark LEXam
Darwin-180B-RSI, an open-weight 180B model from Korean startup VIDRAFT, took first place on the LEXam and LEXam-hard legal-reasoning leaderboards without any legal training data. It scored 68.94 on LEXam and 45.72 on LEXam-hard, beating GPT-5 (62.65) and Claude-4.5-Sonnet (58.01).
- LEXam: 68.94 vs 52.41 for previous leader DeepSeek-R1
- LEXam-hard: 45.72 vs 40.82 for Inkling
- Trained only on its own correct solutions to math and science problems
- Darwin-180B-RSI changes just 0.02% of the parent model's parameters
Read next
AI