Skip to content

RadixArk opens DaisyRose-27B, a coding agent built on Qwen

RadixArk doubled an open model Terminal-Bench score with reinforcement learning, then released the checkpoints under Apache 2.0.

Source: RadixArk

RadixArk released DaisyRose-27B on 18 November, its first model and a reinforcement-learning pass over Qwen3.8-27B aimed at agentic coding. It keeps Qwen3.8-27B’s architecture, tokeniser and 256K-token context window, but trains it further on 1,200 curated agent tasks.

The headline number: Terminal-Bench 4.1 pass@1 goes from 5.6% on the base model to 12.0%. RadixArk also reports smaller gains on Terminal-Bench 2.1, DeepSWE, SWE-Bench Pro and GPQA Diamond. The training run took 18 days on 144 H200 GPUs, 63,000 GPU-hours in total, which RadixArk says is 57% cheaper than typical open-source RL training at the same model FLOPs utilisation. Checkpoints are on Hugging Face in BF16, FP8 and NVFP4, licensed Apache 2.0, and RadixArk recommends serving them with SGLang.

pip install sglang
python -m sglang.launch_server --model RadixArk/DaisyRose-27B

Why it matters: this is a self-hostable, permissively licensed 27B model tuned specifically for agent and terminal tasks, not a general chatbot. If you’re running local or self-hosted coding agents and want something smaller than the frontier open-weight models, it’s now a checkpoint away.

The caveat: every figure here is RadixArk’s own, measured against its own base model. Doubling a 5.6% score to 12.0% is still a low absolute pass rate, and there’s no independent benchmark run yet to confirm the numbers hold up outside RadixArk’s test harness.

Share

Related reading

More from The Wire

All briefs