RadixArk Ships DaisyRose-27B, a 27B Agentic Coding Model Trained With Reinforcement Learning Only

MarkTechPost
Read full story· 5 min
RadixArk released DaisyRose-27B, a 27B-parameter agentic coding model post-trained solely with reinforcement learning from Qwen3.8-27B. It doubles Terminal Bench 4.1 pass@1 score and runs efficiently on a single GPU with open weights under Apache 2.0.

What happened

RadixArk released DaisyRose-27B, a 27-billion-parameter coding model, this week. The model was post-trained from Qwen3.8-27B using only reinforcement learning, without supervised fine-tuning or distillation. DaisyRose-27B is available under an Apache 2.0 license with open weights and is designed to run efficiently on a single GPU. The model achieved more than double the previous score on the Terminal Bench 4.1 coding benchmark after 18 days of training.

Key details

  • Model size: 27.8 billion dense parameters, with a 27-layer vision encoder and 262,144-token context window.
  • Terminal Bench 4.1 pass@1 score increased from 5.6 to 12.0 after training.
  • Training used 144 H200 GPUs for 18 days, totaling 63,000 GPU-hours.
  • Weights are available in BF16, FP8, NVFP4, and GGUF formats on Hugging Face.
  • Runs on a single GPU, requiring about 52 GB (BF16) or 23 GB (NVFP4) of memory.

Why it matters

Developers and organizations can now access an open-weight coding model that offers improved coding benchmark performance and can be deployed on standard hardware. The Apache 2.0 license allows for broad commercial and research use without restrictive terms. The model’s efficiency and performance gains may reduce costs and hardware requirements for teams working on code generation and agentic AI tasks.


More in Machine Learning