RadixArk Releases DaisyRose-27B Model Using New Reinforcement Learning Framework
RadixArk has launched DaisyRose-27B, a model built on Qwen3.8-27B that utilizes a new reinforcement learning framework to significantly improve performance on long-horizon agent tasks.
🔌Chips & HardwarePhoto: Hacker News Front Page
Key points
- DaisyRose-27B doubles Terminal-Bench 4.1 performance compared to its base model.
- The Miles reinforcement learning framework enables asynchronous training to improve GPU utilization.
- Custom sandboxing infrastructure reduced cloud costs by 80% during the nine-week training project.
RadixArk has unveiled DaisyRose-27B, its debut model developed by fine-tuning the Qwen3.8-27B base model. The project utilized a proprietary reinforcement learning framework called Miles, which allowed the team to train the model on 1,200 curated long-horizon tasks over an 18-day period using 144 H200 GPUs. According to the company, this process improved the model's performance on the Terminal-Bench 4.1 benchmark from 5.6% to 12.0%.
The training process was designed for high efficiency, achieving a 57% reduction in training costs compared to standard open-source trainers. By employing an asynchronous training mode, the system allowed the agent to continue working while the model learned, reaching a throughput of 16,800 tokens per second. The team also developed a custom sandboxing service to manage nearly one million task environments, reporting that the infrastructure was 80% cheaper than typical cloud alternatives.
RadixArk is now making the DaisyRose-27B model available in multiple formats, including BF16, FP8, and NVFP4, via Hugging Face. Additionally, the Miles reinforcement learning framework has been released as an open-source tool for other developers to utilize in training their own specialized agents.
Source Hacker News Front Page — radixark.com/blog/daisyrose-27b ↗
Read next
🏢Big TechAnduril CEO Palmer Luckey Calls for Jensen Huang’s Reinstatement at Nvidia
Anduril CEO Palmer Luckey and other industry leaders are pushing for the return of Jensen Huang to Nvidia, arguing his departure weakens U.S. interests. Meanwhile, Nvidia…
🌐Work & SocietyDC Studios pivots strategy under Skydance leadership
Following Skydance’s acquisition of Warner Bros. Discovery, DC Studios is shifting its cinematic strategy toward high-profile heroes, potentially drawing inspiration from…
⚖️Regulation & PolicyAmazon Launches SageMaker JumpStart Model Leaderboard for Standardized AI Comparisons
Amazon has introduced a new leaderboard within SageMaker Studio to help developers compare foundation models based on consistent benchmarks for quality, speed, and cost.
🏢Big TechTake-Two CEO Strauss Zelnick Defends Hands-Off Approach as GTA VI Debuts
Take-Two Interactive CEO Strauss Zelnick leads the launch of Grand Theft Auto VI, a title estimated to cost up to $1.5 billion, while maintaining he does not personally p…
🏢Big TechMicrosoft Unveils Windows Server 2028 with Security-First Default Configuration
Microsoft has launched the public preview of Windows Server 2028, introducing a 'hardened-by-default' security stance aligned with Department of Defense benchmarks.
🤖AI ModelsGoogle Expands Interactive World-Building with Genie 4 Launch
Google has released its latest world-generation model, Genie 4, to premium AI Ultra subscribers, introducing real-time environment modification and audio integration.