Understand this story
The English summary is being updated. See the reporting timeline below.
Latest development · Report
DaisyRose-27B Released, Improving Agent Post-Training Efficiency and Benchmark PerformanceDevelopments
1 confirmed developments
- Source:Hugging Face
Reports and discussion timeline
Browse reports and discussion by time. Their evidence levels remain distinct.
Nov 19
Hugging FaceFirst-partySelected
DaisyRose-27B Released, Improving Agent Post-Training Efficiency and Benchmark Performance
DaisyRose-27B has been released, outperforming the original Qwen3.8-27B on every benchmark listed in the article, including an increase in pass@1 on Terminal Bench 4.1 from 5.6% to 12.0%. The model was trained on approximately 1,200 curated tasks, with the release training run lasting several weeks on 144 H200 GPUs and consuming a total of 63,000 GPU-hours. Its in-house backend achieved 37% MFU, reduced GPU time by 57% compared with open-source trainers, and increased end-to-end training speed by approximately 2×.
Nov 18
Popular on Hacker News
RadixArk Releases DaisyRose-27B, Terminal-Bench 4.1…
RadixArk has released DaisyRose-27B, based on Qwen3.8-27B, raising its Terminal-Bench 4.1 pass@1 score from 5.6% to 12.0%. The model was trained using the Miles reinforcement learning framework on approximately 1,200 curated tasks with 144 H200 GPUs for 18 days, with long-horizon agent sessions reaching up to 1M tokens. RadixArk has also made BF16, FP8, and NVFP4 checkpoints available for download and further development.
AI Trends