RadixArk model lifts an agent test score from 5.6% to 12.0%
RadixArk used only 1,200 tasks, 18 days and 144 GPUs to improve the model.
RadixArk released DaisyRose-27B, its first model, with the open license Apache 2.0. It is a changed version of Qwen3.8-27B, an open model. RadixArk used reinforcement learning, a method that gives the model a score for each task that it completes. The score on Terminal-Bench 4.1, a test of command line work, is up from 5.6% to 12.0%. This is the rate of tasks completed on the first try.
How it works
RadixArk made about 1,200 tasks, and each task is for a mistake that the initial model makes. Each try operates in a sandbox, a clean copy of the task environment, for a maximum of 1.5 hours. The agent continues to work while the model learns.
The numbers
RadixArk tells that:
- its trainer used 57% less GPU time than the open-source trainer.
- the work from start to end was about 2x as fast.
- sandbox cost was about 80% lower than with the usual cloud suppliers.
What is not known
All numbers come from RadixArk. It tells that the model improves on other tests, but gives no results. The model completes only 12.0% of the tasks on the first try.
Insight: One open model now feeds rival projects
DaisyRose-27B is made from Qwen3.8-27B. In September 2026, PrismML released a compressed version of the same Alibaba model, from about 54 gigabytes to below 6. Different firms add layers on one open model, which reads as a platform for firms that work against each other. For leaders: the license of the model and the company for each layer are two different decisions.
Sources
Posted