RadixArk releases DaisyRose-27B for long-running agent tasks

RadixArk released DaisyRose-27B on November 18, an open model built by applying reinforcement learning to Qwen3.8-27B for long-running agent tasks.
The company's release announcement says the model was trained on 1,200 curated tasks over 18 days using 144 NVIDIA H200 GPUs. RadixArk developed the tasks with Eigent AI and used its open-source Miles training framework.
BF16, FP8 and NVFP4 checkpoints are available under the Apache 2.0 license. The model retains the base model's image understanding and 256,000-token context window. Training sessions reached one million tokens through summarization across multiple trajectories; that does not mean the released model has a million-token context window.
RadixArk reports that Terminal-Bench 4.1 pass@1 increased from 5.6% for the base model to 12.0%. Those are company-reported results, not an independent evaluation.
The release gives developers another downloadable, openly licensed model for testing extended coding and tool-use workflows, with multiple numerical formats and a permissive license for adaptation.
Sources and citations
Revision history
Published .