DaisyRose-27B: Pushing the Frontier of Post-Training Efficiency

Hugging Face
Read full story· 6 min
RadixArk released DaisyRose-27B, a fine-tuned version of Qwen3.8-27B trained with reinforcement learning on long-horizon tasks, improving Terminal-Bench 4.1 pass@1 from 5.6% to 12.0%. It supports a 256K-token context window and is available under Apache 2.0.

What happened

RadixArk has launched DaisyRose-27B, a model fine-tuned from Qwen3.8-27B using reinforcement learning on long-horizon agent tasks. The training took 18 days on 144 H200 GPUs, using about 1,200 curated tasks. DaisyRose-27B achieved a Terminal-Bench 4.1 pass@1 score of 12.0%, up from 5.6% for the base model. The model is released under the Apache 2.0 license and supports a 256K-token context window.

Key details

  • DaisyRose-27B was trained with reinforcement learning using the Miles framework.
  • Training involved 1,200 tasks and lasted 18 days on 144 H200 GPUs.
  • Terminal-Bench 4.1 pass@1 improved from 5.6% (Qwen3.8-27B) to 12.0%.
  • The model supports a 256K-token context window and image understanding.
  • It is available under the Apache 2.0 license in multiple checkpoint formats (BF16, FP8, NVFP4).

Why it matters

Developers and organizations working with large language models now have access to a more capable open-source model for long-context and complex reasoning tasks, with improved performance on several benchmarks. The training process used less GPU time and reduced infrastructure costs, making advanced model development more accessible. Users can deploy DaisyRose-27B under a permissive license, enabling broader adoption and integration into commercial and research applications.


More in LLM & Text Generation