
DaisyRose-27B
Visit DaisyRose-27B websiteOpen-weight 27B model post-trained with RL on 1,200 agent tasks, doubling Terminal-Bench pass@1 to 12%
- Made by
- RadixArk
- Listed by
- Jin
- Pricing
- Open source
- Added
- 18 Nov 2026
- Our coverage
- Read our story
About DaisyRose-27B
RadixArk's first open model, built on Qwen3.8-27B using their Miles RL framework. Trained for 18 days on 144 H200 GPUs with 1,200 curated long-horizon agent tasks spanning ML, software engineering, and science. Achieves 12.0% pass@1 on Terminal-Bench 4.1 (vs 5.6% base). Released in BF16, FP8 and NVFP4 formats under Apache 2.0. Runs on SGLang, vLLM and standard inference stacks.
