parallelquant
November 20, 2026 · MarkTechPost

RadixArk releases open DaisyRose-27B coding agent trained only with RL

RadixArk released DaisyRose-27B, an Apache 2.0 agentic coding model post-trained from Qwen3.8-27B using only reinforcement learning (RL). It more than doubles the base model's Terminal Bench 4.1 pass@1 score to 12.0 and runs on a single GPU.

Why it matters: Skipping supervised fine-tuning and relying purely on RL shows that open labs can extract large agentic gains from existing base models without costly curated demonstrations. The absolute score remains low, but a permissively licensed, single-GPU coding agent lowers the barrier for local and private deployments, and reinforces how Qwen bases have become the default foundation for open-model work.

Related updates