DaisyRose-27B: Pushing the Frontier of Post-Training Efficiency | RadixArk Blog
RadixArk has released DaisyRose-27B, its first reinforcement-learning-trained model built on Qwen3.8-27B. Trained on 1,200 curated long-horizon agent tasks with sessions up to one million tokens using the Miles framework, the model improves Terminal-Bench 4.1 pass@1 from 5.6% to 12.0%. The company achieved a 57% lower training cost and 80% lower sandbox infrastructure costs by utilizing asynchronous training and dynamic sandbox memory allocation. Open article →