RadixArk Releases DaisyRose-27B: Post-Training Efficiency Breakthrough
RadixArk introduces DaisyRose-27B, built on Qwen3.8-27B and trained using reinforcement learning via Miles with high efficiency.

RadixArk has unveiled DaisyRose-27B, its first model built upon the foundation of Qwen3.8-27B. Trained using reinforcement learning with the Miles framework, the release focuses on improving post-training efficiency for long-horizon agent tasks. The project was completed using 1,200 curated tasks over an 18-day period on 144 H200 GPUs.
DaisyRose-27B retains the original Qwen3.8-27B architecture, tokenizer, chat template, image understanding features, and its 256K-token context window. The model is released under the Apache 2.0 license in BF16, FP8, and NVFP4 formats. By default, the model thinks before answering, returning its reasoning in a dedicated field and the final answer separately.
Benchmark Performance and Improvements
Evaluations show that DaisyRose-27B outperforms the original Qwen3.8-27B across every tested benchmark. The most notable improvement occurs in Terminal-Bench 4.1, where the pass@1 rate jumps from 5.6% to 12.0%.
Additional performance gains are observed across coding, reasoning, and instruction-following tests:
- Terminal-Bench 2.1 rises from 73.0% to 75.5%
- DeepSWE v1.1 increases from 42.2% to 46.2%
- SWE Bench Pro v2 moves from 91.3% to 94.7%
- HLE (text-only) shifts from 33.9% to 34.3%
- IFBench improves from 79.5% to 82.4%
- GPQA Diamond goes from 89.2% to 90.7%
RadixArk achieved these outcomes through a combination of carefully curated training data built in partnership with Eigent AI, an efficient asynchronous training backend called Miles, and a robust sandbox service based on AgentENV.
What it means for developers
Developers working with complex, long-horizon agent tasks can leverage DaisyRose-27B for better multi-step reasoning, coding, and planning capabilities. For inference, RadixArk recommends serving the model using SGLang and DSpark speculative decoding. Developers can also explore various top AI models cheaply through one API at https://apixoai.online.
Furthermore, the training infrastructure behind DaisyRose-27B highlights advancements in resource utilization. The training backend achieved a 57% lower training cost compared to open-source alternatives, maintaining stable operations over 63,000 GPU-hours.
Source: DaisyRose-27B: Pushing the Frontier of Post-Training Efficiency — Hugging Face blog. Written by the Apixo team from that report.
One key for Claude, GPT, GLM, DeepSeek and more. Pay per token with crypto.
Get your API key

