Wink Pings

27B Parameter Open-Source Programming Model Doubles Performance, Runs on a Single GPU: Post-Training Entirely via Reinforcement Learning

RadixArk has launched DaisyRose-27B, an agentic programming model built on Qwen3.8-27B. Its post-training phase was optimized entirely using reinforcement learning, achieving a pass@1 score of 12.0 on Terminal Bench 4.1—more than double that of its base model. It supports single-GPU deployment and is released under the Apache 2.0 open-source license. The news was first published by tech media MarkTechPost.

RadixArk has officially released DaisyRose-27B, an agentic programming model. The report comes from tech media MarkTechPost, written by Asif Razzaq.

This 27B-parameter model was post-trained based on Qwen3.8-27B. Unlike the multi-stage training workflows common in the industry, its entire post-training process only uses reinforcement learning—no other optimization methods like supervised fine-tuning were incorporated.

In terms of performance metrics, DaisyRose-27B achieved a pass@1 score of 12.0 in the Terminal Bench 4.1 benchmark, more than double that of its base model.

For readers without a technical background, two basic concepts: Terminal Bench is a benchmark suite that simulates real-world terminal development scenarios, covering common tasks such as script writing, tool invocation, and environment debugging; pass@1 refers to the probability that the model's first generated output successfully completes the task, making it one of the most critical performance metrics for programming models.

Deployment accessibility is another major highlight of this model. Currently, large models of the same parameter size mostly require multiple GPUs in parallel for inference. DaisyRose-27B has undergone inference-side optimizations, enabling the full runtime process to run on a single GPU, so individual developers and small teams can deploy and use it locally without setting up a dedicated computing cluster.

Regarding open-source licensing, DaisyRose-27B releases its full model weights under the Apache 2.0 license. Individuals and enterprises can freely modify, redistribute, and even use it for commercial purposes without requiring additional authorization or facing commercial royalty restrictions.

发布时间: 2026-11-21 01:42