DaisyRose-27B achieves higher benchmarks with efficient post-training
RadixArk releases DaisyRose-27B, a model that outperforms Qwen3.8-27B on key benchmarks using a new asynchronous training framework and cost-effective sandbox infrastructure.
RadixArk has released DaisyRose-27B, a new open-weight language model that demonstrates significant improvements in post-training efficiency and performance. Published in November 2026, this release highlights how careful data curation and specialized infrastructure can enhance model capabilities without proportional increases in computational cost. The model is built upon the Qwen3.8-27B architecture but utilizes a novel training approach to correct specific weaknesses found in the base model.
The release marks a shift toward more targeted reinforcement learning strategies, where training data is generated specifically to address known model failures. By combining high-quality, error-focused tasks with an asynchronous training backend, RadixArk achieved substantial gains in both benchmark scores and operational efficiency. This approach allows the model to learn from long-horizon agent tasks while maintaining stable training dynamics over weeks of continuous operation.
What happened
DaisyRose-27B outperforms the original Qwen3.8-27B model across every evaluated benchmark. The most notable improvement appears on Terminal Bench 4.1, where the pass@1 score increased from 5.6% to 12.0%. These results were derived from the higher of BF16 and FP8 runs for each benchmark, with differences between the two precision formats remaining under one point wherever both were tested. The model card provides full details on comparison sources and evaluation settings to ensure reproducibility.
The development process relied on three core components: carefully designed training tasks, an efficient training backend called Miles, and a robust sandbox service. RadixArk partnered with Eigent AI to create the training data, starting from a large pool of tasks in fields such as machine learning, software engineering, and science. From thousands of candidates, approximately 1,200 tasks were selected for the final training set based on strict criteria aimed at maximizing learning utility.
Each task was designed to target specific mistakes made by the original Qwen3.8-27B model. The team ran the base model on every candidate task and redesigned them around the errors they wanted to eliminate. Difficulty levels were adjusted until the solve rate fell within a useful range for learning. Additionally, each task requires multiple interdependent steps, forcing the agent to reason, plan, and execute over long jobs. Attempts can last up to 1.5 hours in a sandbox, involving more than 100 tool calls and generating over 256,000 output tokens. Tasks that could be solved without addressing the intended problem or those with design flaws were removed or fixed.
How it works
The training infrastructure uses Miles, an open-source reinforcement learning framework developed by RadixArk, operating in fully asynchronous mode. In this setup, the agent continues working on tasks while the model learns from previous trajectories, ensuring neither process waits for the other. The system allocates 128 H200 GPUs for inference on agent tasks and 16 additional GPUs for the trainer. After every training step, the updated model is immediately distributed to the 128 GPUs running the agents.
This backend is optimized for agent training, where each step learns from a handful of very long sessions, some reaching up to 1 million tokens. These sessions are folded into multiple trajectories via harness summarization. On this release, 37% of each GPU’s peak compute was utilized for useful training work, known as model FLOPs utilization (MFU). In contrast, the open-source trainer built on Megatron achieved only 16% MFU on comparable runs. This efficiency gain translates to 57% less GPU time required for the same training outcome.
The efficiency advantage is even more pronounced for mixture-of-experts models. For Qwen3.6-35B-A3B, the speedup is 6.2×, and for the roughly 300-billion-parameter GLM-5.3-Flash, it exceeds 10×. End-to-end, the release run trained at about 16,800 tokens per second across its 144 GPUs, compared to 8,300 tokens per second for the open-source trainer. The system remained stable over weeks of training, accumulating 63,000 GPU-hours without degradation in speed or efficiency.
Key details
- DaisyRose-27B improves pass@1 on Terminal Bench 4.1 from 5.6% to 12.0% compared to Qwen3.8-27B.
- The training set consists of 1,200 tasks selected from thousands of candidates, each targeting specific base model errors.
- Miles, the asynchronous training framework, achieved 37% MFU compared to 16% for the open-source Megatron-based trainer.
- The system used 144 H200 GPUs, with 128 for inference and 16 for training, processing up to 16,800 tokens per second.
- Sandbox infrastructure costs were 80% lower than popular cloud providers, with 99.95% creation success rate since late October.
- Training involved sessions up to 1.5 hours long, with over 100 tool calls and 256K output tokens per attempt.
Why it matters
For engineers building AI agents, this release demonstrates that post-training efficiency can be dramatically improved through asynchronous architectures. Traditional synchronous training often bottlenecks progress because the model must wait for data generation or vice versa. By decoupling these processes, RadixArk shows that it is possible to maintain high GPU utilization while handling long-horizon tasks. This is critical for developers who need to train models on complex, multi-step problems without incurring prohibitive computational costs.
The focus on error-specific data curation also offers a practical lesson in model improvement. Instead of generic pre-training on vast datasets, targeting known failure modes with high-quality, difficult tasks can yield significant performance gains. This approach requires rigorous validation to ensure tasks are not solvable via shortcuts, but the resulting models are more robust in real-world applications. The transparency in sharing the Miles framework and AgentENV contributions further enables the community to adopt these efficiency techniques.
What you can do
- Download DaisyRose-27B from the Hugging Face hub to test its performance on your specific workflows.
- Review the SGLang serving guide to deploy the model efficiently in your production environment.
- Use the Miles framework to train your own agents, leveraging the asynchronous training methodology described.
- Evaluate the model on your internal benchmarks and share results in the model’s Community tab to help identify remaining shortcomings.
- Study the task selection criteria used by RadixArk and Eigent AI to improve your own data curation pipelines.
- Explore the AgentENV project to build cost-effective sandbox environments for your own agent training needs.



