# RadixArk releases DaisyRose-27B to showcase its agent training infrastructure

> Source: <https://runtimewire.com/article/radixark-doubles-a-qwen-model-s-terminal-score-with-1-200-tasks-and-144-gpus>
> Published: 2026-11-18 20:02:05+00:00

**Recommendation:** Budget for long RL post-training runs on agent benchmarks: RadixArk's 18-day run on 144 H200 GPUs took Qwen3.8-27B from 5.6% to 12.0% on Terminal-Bench 4.1.
More in the playbook "Fine-tuning and data": <https://wpnews.pro/playbooks/training-data>

# RadixArk releases DaisyRose-27B to showcase its agent training infrastructure

**Co-founders Ying Sheng and Banghua Zhu released DaisyRose-27B on November 18th, an Apache 2.0 post-train of Qwen3.8-27B built with RadixArk's Miles framework.**

        By [RuntimeWire Staff](https://runtimewire.com/author/runtimewire-staff)
        · Published 
        · Updated 

RadixArk co-founders Ying Sheng and Banghua Zhu have shipped the company's first model, DaisyRose-27B, a reinforcement-learning post-train of Qwen3.8-27B that RadixArk says more than doubled the base model's Terminal-Bench 4.1 pass@1 score, from 5.6% to 12.0%, after 18 days on 144 Nvidia H200 GPUs.…
