RadixArk has released DaisyRose-27B, a new model built upon Qwen3.8-27B. This model was trained using reinforcement learning with a focus on long-horizon agent tasks, utilizing sessions up to 1 million tokens. DaisyRose-27B demonstrates significant improvements on benchmarks like Terminal-Bench 4.1, increasing its pass@1 score from 5.6% to 12.0%, and also shows gains in reasoning, instruction following, and coding tasks. The training process involved curated tasks developed in partnership with Eigent AI and utilized RadixArk's open-source reinforcement learning framework, Miles. AI
IMPACT Enhances agent task performance and efficiency in post-training, potentially improving reasoning and coding capabilities.
RANK_REASON New model release from a lab (RadixArk) with performance benchmarks. [lever_c_demoted from frontier_release: ic=1 ai=1.0]
- Apache Software License 2.0
- DaisyRose-27B
- DSpark
- Eigent AI
- .openai
- Qwen3.8-27B
- RadixArk
- SGLang
- Terminal-Bench 4.1
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →