RadixArk developed DaisyRose-27B, its first open-weight model, by applying reinforcement learning to Qwen3.8-27B over 18 days on 144 H200 GPUs. The training used 1,200 curated long-horizon agent tasks and increased the model’s Terminal-Bench 4.1 score to 12.0% from 5.6%.
The model allows agent sessions of up to 1M tokens and has day-0 support for deployment via SGLang.
