DaisyRose-27B: A Leap Forward in Post-Training Efficiency
RadixArk has unveiled DaisyRose-27B, a significant advancement in the field of large language model development. This new model, built upon the Qwen3.8-27B architecture, demonstrates a remarkable improvement in post-training efficiency. By leveraging reinforcement learning and meticulously curated datasets, DaisyRose-27B achieves substantial gains in performance while drastically reducing training costs and time. This breakthrough pushes the frontier of what’s possible in efficient AI model development.
Table of Contents
The model’s success is highlighted by its performance on the Terminal-Bench 4.1 benchmark, where it improved the pass@1 score from 5.6% to 12.0%. Beyond this, DaisyRose-27B shows enhanced capabilities across various reasoning, instruction-following, and coding tasks, underscoring its versatility and effectiveness.
Architectural Innovations and Performance Metrics
DaisyRose-27B retains the core architecture, tokenizer, and chat template of its predecessor, Qwen3.8-27B. Crucially, it inherits the impressive 256K-token context window, enabling it to process and reason over extensive sequences of information. The training process utilized sessions up to 1 million tokens, employing harness summarization to effectively span multiple long trajectories.
The model’s training was conducted on 144 H200 GPUs over 18 days, focusing on 1,200 carefully curated tasks. This approach resulted in a 57% reduction in training cost compared to standard open-source trainers. Furthermore, the end-to-end training process was accelerated by 2x, achieving approximately 16,800 tokens per second across all GPUs.
Engineering Trade-offs and Training Backend Efficiency
The remarkable post-training efficiency of DaisyRose-27B is attributed to three key components: high-quality training data, an efficient training backend, and a robust sandbox service. The training data was developed in partnership with Eigent AI, focusing on tasks that specifically targeted Qwen3.8-27B’s weaknesses and required multi-step reasoning.
RadixArk’s proprietary reinforcement learning framework, Miles, played a pivotal role. Operating in a fully asynchronous mode, Miles ensures that the agent tasks and model learning proceed concurrently, maximizing GPU utilization. This optimized backend achieved 37% of peak compute for useful training work, a significant improvement over the 16% typically seen with open-source trainers like Megatron on comparable runs.
Field Readiness and Verdict on DaisyRose-27B
DaisyRose-27B is readily available for use, with checkpoints released under the Apache 2.0 license. For optimal performance, serving the model with SGLang and DSpark speculative decoding is recommended. The SGLang cookbook offers configurations for various GPU setups.
The stability of the training process over weeks, with consistent speed and efficiency, demonstrates the robustness of RadixArk’s infrastructure. The sandbox service, built on the open-source AgentENV project, also proved highly reliable, achieving over 99.95% success rate for sandbox creations in recent operations and significantly lower cloud infrastructure costs. This focus on efficiency and reliability makes DaisyRose-27B a compelling option for researchers and developers looking to push the boundaries of AI without prohibitive costs, akin to how innovations in areas like Linux in browser are making complex software more accessible.
Adapted from the original RadixArk article. Source: Hugging Face Technical Blog


