RadixArk Ships DaisyRose-27B, a 27B Agentic Coding Model Trained With Reinforcement Learning Only
Summary
- RadixArk has released DaisyRose-27B, a 27.8B parameter agentic coding model trained purely via reinforcement learning without SFT or distillation.
- The model was trained over 18 days using 144 H200 GPUs on RadixArk's Miles framework, achieving 37% MFU and a 2x throughput speedup.
- On Terminal Bench 4.1, the pass@1 score more than doubled from 5.6 to 12.0 compared to its Qwen3.8-27B base model.
- DaisyRose-27B features a 262,144-token context window, supports text and vision inputs, and is available under the Apache 2.0 license.
Why it matters
DaisyRose-27B demonstrates that pure reinforcement learning can significantly boost coding and agentic capabilities in a dense 27B model, providing an efficient open-weight option that runs on a single GPU.