DaisyRose-27B: 1,200-task RL took Qwen3.8 from 5.6% to 12% on Terminal-Bench 4.1 (www.radixark.com)

🤖 AI Summary
RadixArk has announced the release of DaisyRose-27B, a reinforcement learning model built on Qwen3.8-27B, which significantly improves its performance on the challenging Terminal-Bench 4.1, boosting the pass rate from 5.6% to 12.0%. This model, trained over 18 days on 144 H200 GPUs, utilized a carefully curated set of 1,200 tasks, focusing on long-horizon agent tasks that require complex reasoning, planning, and execution. By implementing a highly efficient training backend and robust sandboxing environment, RadixArk achieved breakthroughs in model performance while managing to lower training costs by 57% compared to traditional methods. The significance of DaisyRose-27B lies in its marked advancements in reasoning and task execution, highlighting the impact of high-quality training data and innovative infrastructure on AI capabilities. The integration of open-source techniques, including the Miles reinforcement learning framework, allowed for asynchronous training and improved GPU utilization. Moreover, the project's cloud infrastructure cost for running sandboxes was reduced by 80%. This release not only sets a new benchmark in model training efficiency but also provides a robust framework for future AI endeavors, encouraging other teams to explore similar approaches for training specialized models.
Loading comments...
loading comments...