What happened
RadixArk released model weights for DaisyRose-27B on 18 November. It built the model from Qwen3.8-27B and used reinforcement learning on tasks with multiple steps.
Why it matters
The results show how performance changed after RadixArk trained the base model on more tasks. The size of the improvement varied by test.
The data
Each pair compares the original Qwen3.8-27B model with DaisyRose-27B on the same benchmark.
| Item | Value | Unit | Date |
|---|---|---|---|
| Terminal-Bench 4.1 · Qwen | 5.6 | percent | 2026-11-18 |
| Terminal-Bench 4.1 · DaisyRose | 12.0 | percent | 2026-11-18 |
| Terminal-Bench 2.1 · Qwen | 73.0 | percent | 2026-11-18 |
| Terminal-Bench 2.1 · DaisyRose | 75.5 | percent | 2026-11-18 |
| DeepSWE v1.1 · Qwen | 42.2 | percent | 2026-11-18 |
| DeepSWE v1.1 · DaisyRose | 46.2 | percent | 2026-11-18 |