DaisyRose-27B: 1,200-task RL took Qwen3.8 from 5.6% to 12% on Terminal-Bench 4.1
AI · Studies · 2026-11-18 19:39
RadixArk released DaisyRose-27B, a model based on Qwen3.8 that improved Terminal-Bench 4.1 scores from 5.6% to 12% through reinforcement learning across 1,200 tasks.