DaisyRose-27B Model Released with Enhanced Post-Training Efficiency
RadixArk has launched DaisyRose-27B, its first model built on Qwen3.8-27B, which has been trained using reinforcement learning on long-horizon tasks. This model shows significant improvements in performance metrics, particularly on the Terminal-Bench 4.1 benchmark, where its pass rate increased from 5.6% to 12.0%. The training process utilized an efficient backend and a robust sandbox service, resulting in lower costs and faster training times.
2 min read1 source
Key takeaways
30-second read- 1Developers can leverage DaisyRose-27B's improved reasoning and instruction-following capabilities for more complex applications in AI, enhancing user interactions.
- 2The model's efficient training methods could inform developers on optimizing their own training processes, potentially reducing costs and time.
- 3With the open-source nature of DaisyRose-27B, builders can customize and adapt the model for specific tasks, fostering innovation in AI applications.
Confirmed
Official source: HuggingFace Blog
Reported
No secondary reporting linked
Verify yourself
Eligibility, local prices and dates for your car, state or utility

RadixArk has introduced DaisyRose-27B, a new AI model that builds on the architecture of Qwen3.8-27B, specifically designed to enhance post-training efficiency. This model has been developed using reinforcement learning techniques and is particularly focused on long-horizon agent tasks, which involve complex reasoning and planning over extended interactions. Notably, DaisyRose-27B has demonstrated a substantial improvement in its performance metrics, especially on the Terminal-Bench 4.1 benchmark, where its pass rate has increased from 5.6% to 12.0%. This advancement suggests a significant leap in the model's ability to handle intricate tasks effectively.
What happened
The training of DaisyRose-27B was accomplished in just 18 days using 144 NVIDIA H200 GPUs, processing a total of 63,000 GPU hours. The training involved 1,200 carefully curated tasks that were designed to target specific weaknesses identified in the original Qwen3.8-27B model. The training sessions were extensive, with some reaching up to 1 million tokens, allowing the model to learn from a wide range of scenarios. The training backend was optimized to allow asynchronous operation, meaning that the model could learn and the agent could execute tasks simultaneously, leading to a more efficient training process.
Why it matters
- Enhanced Capabilities: Developers can utilize DaisyRose-27B's improved reasoning and instruction-following capabilities for applications requiring complex decision-making.
- Cost Efficiency: The model's training process is reported to be 57% more cost-effective in terms of GPU time compared to previous methods, which could significantly lower the barrier for developers working on similar projects.
- Open Source Access: As DaisyRose-27B is available under the Apache 2.0 license, builders can modify and adapt the model for their specific needs, promoting a culture of innovation and collaboration in the AI community.
Context and caveats
The development of DaisyRose-27B involved collaboration with Eigent AI to create high-quality training data. The model's architecture retains key features from Qwen3.8-27B, including its tokenizer and chat template, ensuring compatibility with existing systems. However, the full implications of its performance improvements will require further testing and validation in real-world applications.
What to watch next
As DaisyRose-27B becomes available for public use, it will be important to monitor how developers implement this model in various applications. Additionally, feedback from the community regarding its performance on different tasks will be crucial for future iterations and enhancements. The ongoing development of open-source tools like Miles and SGLang will also be significant in shaping the landscape of AI model training and deployment.
Sources
1 official- [1]DaisyRose-27B: Pushing the Frontier of Post-Training EfficiencyOfficial
HuggingFace Blog · huggingface.co · published Nov 18
Newsletter
Weekly AI signal. No spam.

Comments
Sign in to join the discussion
Loading comments…