RadixArk released DaisyRose-27B, a new open-source model built on the Qwen3.8-27B architecture. The project trained the model using the open-source Miles reinforcement learning framework and the AgentENV sandbox service. DaisyRose-27B supports a 256K-token context window and includes image understanding capabilities.
The model required 18 days of training on 144 H200 GPUs to process 1,200 curated tasks. RadixArk processed 16,800 tokens per second during the release run and utilized 944,000 sandboxes on the AgentENV service. The setup achieved 57% lower GPU time and 37% model FLOPs utilization compared to open-source baselines.
Developers can use DaisyRose-27B with SGLang and DSpark for speculative decoding. The model improved the Terminal-Bench 4.1 pass@1 score from 5.6% to 12.0% in testing. This release offers a resource-efficient alternative for building autonomous agents that require strong coding and terminal interaction skills.


