| |
DaisyRose-27B: 1,200-task RL took Qwen3.8 from 5.6% to 12% on Terminal-Bench 4.1
RadixArk released DaisyRose-27B, a 27B parameter model built on Qwen3.8 and trained with reinforcement learning on 1,200 curated long-horizon tasks, improving Terminal-Bench 4.1 performance from 5.6% to 12.0% pass@1. The model was trained in 18 days on 144 H100 GPUs using high-quality task design, efficient asynchronous training infrastructure, and robust sandboxes to prevent reward hacking. DaisyRose-27B shows improvements across multiple benchmarks including reasoning, instruction following, and coding tasks, with the team releasing multiple checkpoint formats for public use.
Read Full Article →
← More Tech news