DaisyRose-27B: Pushing the Frontier of Post-Training Efficiency
Source Entity
Hugging Face - Blog

RadixArk has released DaisyRose-27B, an AI model built on Qwen3.8-27B optimized for long-horizon agent tasks. The model significantly improves performance on complex benchmarks using efficient reinforcement learning techniques.
The Emergence of DaisyRose-27B: A Milestone in Agentic AI
RadixArk’s release of DaisyRose-27B marks a significant evolution in the development of agentic large language models (LLMs). By leveraging the foundation of the Qwen3.8-27B architecture, the team has introduced a model specifically fine-tuned for long-horizon agent tasks. This development is pivotal because it addresses a critical bottleneck in contemporary AI: the ability to maintain coherence and reasoning over extended sequences of interaction, effectively pushing the frontier of post-training efficiency.
Technical Architecture and Efficiency
The training methodology behind DaisyRose-27B is particularly noteworthy for its resource efficiency. By utilizing reinforcement learning via the 'Miles' framework, the team achieved substantial performance improvements using only 1,200 curated tasks over a span of 18 days on 144 H200 GPUs. This demonstrates that high-quality, targeted data curation—rather than mere scale—is becoming the primary driver of performance gains in modern model development, a trend that is likely to define the next generation of open-weights models.
Benchmarking Success and Reasoning Gains
The impact of this training is clearly visible in the model’s performance metrics. Most notably, DaisyRose-27B more than doubled the pass rate on Terminal-Bench 4.1, jumping from 5.6% to 12.0%. This improvement in terminal-based reasoning suggests that the model is significantly better at executing multi-step command-line operations, a prerequisite for autonomous software engineering agents. Beyond this specific benchmark, the model shows generalized improvements in instruction following and coding, validating its utility as a versatile tool.
Long-Horizon Context and Trajectory Management
One of the most impressive technical feats of DaisyRose-27B is its handling of sessions up to 1 million tokens. By utilizing 'harness summarization' to span multiple trajectories, the model manages to maintain its reasoning capabilities across massive context windows. By retaining the core 256K-token context window and image understanding capabilities of the original Qwen3.8-27B, the model ensures that it does not lose its multimodal strengths while gaining specialized agentic prowess.
Broader Implications for Open Source AI
The decision to release the checkpoints under the Apache 2.0 license is a major boost for the open-source community. It allows researchers and developers to inspect the model's properties, adjust inference settings, and integrate it into their own agentic workflows. As companies continue to compete for dominance in the 'agentic' space, the availability of high-performing, efficiently trained models like DaisyRose-27B lowers the barrier to entry for building complex, autonomous AI systems.
Future Trends in Agentic Training
Looking ahead, the success of the Miles reinforcement learning approach suggests that we are moving toward a future where AI agents are trained on 'quality over quantity.' The ability to perform complex tasks with relatively small, curated datasets indicates that future models will be more specialized and computationally efficient. RadixArk has set a new precedent for how to effectively adapt existing base models for long-horizon tasks, providing a roadmap for others to follow in the pursuit of more reliable, task-oriented artificial intelligence.