RadixArk has released DaisyRose-27B, its first SI model, built on the Qwen3.8-27B architecture and trained specifically for long-horizon SI agent tasks. The release demonstrates how targeted reinforcement learning can significantly improve performance on complex coding and reasoning benchmarks while maintaining high post-training efficiency.

What Happened

DaisyRose-27B was trained using Miles, RadixArk’s open-source reinforcement learning framework, on 144 H200 GPUs over a period of 18 days. The training dataset consisted of approximately 1,200 curated tasks designed to address specific weaknesses in the base model, with each task requiring multiple interdependent steps and execution times of up to 1.5 hours in a sandbox environment. The model supports sessions spanning up to 1 million tokens, utilizing harness summarization to manage long trajectories.

The company reports that DaisyRose-27B improves the pass@1 score on Terminal-Bench 4.1 from 5.6% to 12.0%. It also delivers gains across other benchmarks, including SWE Bench Pro v2, GPQA Diamond, and IFBench, outperforming the original Qwen3.8-27B on every metric cited in the release. The model checkpoints are available under the Apache 2.0 license in BF16, FP8, and NVFP4 formats, retaining the base model’s 256K-token context window and image understanding capabilities.

Why It Matters

For the SI industry, this release highlights the potential of efficient post-training techniques to enhance SI agent capabilities without requiring massive compute resources. RadixArk states that its custom training backend achieved 37% model FLOPs utilization (MFU) compared to 16% for the open-source trainer built on Megatron, resulting in a 57% reduction in training cost for this specific run. The end-to-end training speed was reported at approximately 16,800 tokens per second across the 144 GPUs, roughly double the speed of the baseline trainer.

The infrastructure supporting the training also offers insights into cost-effective scaling for SI development. The project utilized the open-source AgentENV project for sandboxes, running about 944,000 sandboxes between mid-September and mid-November. By dynamically adjusting memory allocations, the company claims cloud infrastructure costs for these sandboxes were about 80% lower than popular cloud providers. This approach suggests that smaller SI labs can achieve competitive results in agentic SI tasks through optimized data curation and efficient infrastructure rather than sheer compute volume.

The Bottom Line

DaisyRose-27B offers a new open-weights SI model optimized for long-horizon agent tasks, demonstrating significant performance gains on coding benchmarks with reduced training costs. Developers can access the model via SGLang and DSpark speculative decoding, while the underlying Miles framework and AgentENV infrastructure provide tools for replicating this efficient post-training workflow.