Urgent.News

What's breaking now, across thousands of outlets.

AI

DaisyRose-27B: Pushing the Frontier of Post-Training Efficiency

RadixArk has unveiled DaisyRose-27B, their latest AI model that pushes the boundaries of post-training efficiency. Built on Qwen3.8-27B, this model underwent reinforcement learning training using Miles, a powerful open-source framework. Trained on long-horizon agent tasks that span up to 1 million tokens, the process was completed in just 18 days using 144 H200 GPUs.

DaisyRose-27B boasts impressive improvements over its predecessor, with an increase in Terminal-Bench 4.1 pass@1 from 5.6% to 12.0%. The gains extend beyond this benchmark, showing positive results across various reasoning, instruction-following, and coding benchmarks. The model maintains Qwen3.8-27B's architecture, tokenizer, chat template, image understanding, and a 256K-token context window.

RadixArk has made the model checkpoints available under the Apache 2.0 license, ensuring accessibility for the wider community. For optimal performance, it is recommended to serve DaisyRose-27B with SGLang and DSpark speculative decoding. Detailed model properties and inference settings can be found in the provided model card.

The training process utilized harness summarization to effectively span multiple trajectories, resulting in 57% lower training cost compared to traditional methods. The open-source trainer, built on Megatron, achieved 16% utilization of the model's FLOPs on comparable runs. This efficiency translates to a 6.2× gain on Qwen3.6-35B-A3B and over 10× improvement on the roughly 300B-parameter GLM-5.3-Flash models.

End-to-end training time was remarkably fast, with the model processing about 16,800 tokens per second across all 144 GPUs. The release run trained for approximately 63,000 GPU-hours without any significant fluctuations, ensuring consistent performance throughout the entire process. During this period, up to 1,566 sandboxes were running simultaneously, with a near-perfect success rate of 99.95% in creating sandboxes for this project.

RadixArk extends gratitude to Eigent AI for their collaboration in training data, the Qwen team for creating Qwen3.8-27B, as well as the dedicated open-source communities behind Miles, SGLang, slime, AgentENV, and Harbor. Feedback on DaisyRose-27B's performance and potential areas of improvement are welcome, as the community continues to explore its capabilities and applications.

Written by urgent.news from Hugging Face's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.

Read the original at huggingface.co →

More in AI

More from Wednesday 18 November →