DaisyRose-27B Achieves Major Efficiency Gains in AI Training
RadixArk introduces DaisyRose-27B, a model trained on Qwen3.8-27B using reinforcement learning with Miles, achieving a 12.0% pass@1 rate on Terminal-Bench 4.1—double the original model’s 5.6%. The model’s success stems from high-quality, curated tasks targeting specific weaknesses, long-horizon training sessions, and sandbox environments preventing reward hacking. Training was optimized via an asynchronous backend, reducing GPU costs by 57% and doubling training speed compared to open-source alternatives. Sandboxing infrastructure, scaled from single to 12 machines, cut cloud costs by 80% while maintaining 99.95% reliability. The model’s checkpoints and training framework are open-source, enabling broader adoption for specialized AI development.
Read the full article
Continue reading at Hacker News →This is an AI-generated summary. Read the original for the full story.