DaisyRose-27B: 1,200-task RL took Qwen3.8 from 5.6% to 12% on Terminal-Bench 4.1
Comments
Ranked by a public rubric, not by votes: what a comment adds counts, not who wrote it. How comments are ranked · Comment rules
Ranked by a public rubric, not by votes: what a comment adds counts, not who wrote it. How comments are ranked · Comment rules
Summary
RadixArk released DaisyRose-27B, a version of Alibaba's Qwen3.8-27B fine-tuned with reinforcement learning on 1,200 curated long-horizon agent tasks over 18 days on 144 H200 GPUs.