Grafy News
Try Grafy free

RadixArk releases DaisyRose-27B agentic coding model

RadixArk has launched DaisyRose-27B, a new coding model trained exclusively via reinforcement learning to significantly boost terminal performance.

Illustration for the news story: RadixArk releases DaisyRose-27B agentic coding modelSource: MarkTechPost

RadixArk has launched DaisyRose-27B, a new 27-billion parameter model specifically designed for agentic coding tasks. This release marks a shift in training methodology, as the model was post-trained from Qwen3.8-27B using reinforcement learning alone, according to MarkTechPost.

What makes DaisyRose-27B different from other coding models?

Most large language models undergo extensive supervised fine-tuning (SFT) using large datasets of human-written code before moving to other stages. However, RadixArk utilized a different approach for DaisyRose-27B. The model was post-trained from the Qwen3.8-27B base model using only reinforcement learning.

By focusing on reinforcement learning, the model is trained to optimize for specific outcomes and success metrics rather than simply mimicking human patterns. This method aims to enhance the model's ability to act as an "agent"—a system capable of taking autonomous steps to solve complex programming problems, such as navigating a file system or debugging errors.

How much does the model improve coding performance?

According to MarkTechPost, DaisyRose-27B shows a significant leap in specialized benchmarks. Specifically, the model more than doubled its performance on the Terminal Bench 4.1 metric, achieving a pass@1 score of 12.0.

Terminal Bench is a critical metric for agentic models because it measures how effectively an AI can interact with a command-line interface to execute tasks. A higher pass@1 score indicates that the model is more likely to successfully complete a terminal-based coding task on its first attempt without requiring multiple corrections or human intervention.

Can DaisyRose-27B run on standard hardware?

One of the practical advantages of the DaisyRose-27B model is its efficiency regarding hardware requirements. Despite its specialized training and agentic capabilities, the model is designed to run on a single GPU.

This accessibility is important for developers and researchers who may not have access to massive industrial computing clusters. Being able to deploy a 27B parameter model on a single unit allows for faster iteration, easier local deployment, and lower costs for integrating agentic coding capabilities into existing workflows.

What is the licensing and origin of the model?

DaisyRose-27B is released under the Apache 2.0 license, which is a permissive open-source license. This allows developers and companies to use, modify, and distribute the model with relatively few restrictions, fostering wider adoption within the developer community.

The model is built upon the foundations of Qwen3.8-27B, a model from the Qwen series, which RadixArk then refined through its reinforcement learning process to specialize in the domain of software engineering and terminal interaction.

How can developers apply agentic coding models?

Agentic models like DaisyRose-27B are intended to go beyond simple code completion. They are designed to function as autonomous collaborators that can understand a high-level instruction, plan the necessary steps, execute commands in a terminal, check the results, and correct their own mistakes.

For developers, this means moving toward a workflow where the AI does not just suggest a line of code, but can actually run tests, install necessary dependencies, and manage complex refactoring tasks across multiple files.

How to explore agentic coding in Grafy

While DaisyRose-27B is a specific model release, the concept of agentic coding—where the AI can run and test its own work—is a core principle of modern AI development environments.

Developers looking to experience an integrated environment where AI assists with the full lifecycle of programming can use Grafy Code. This AI coding studio is designed to handle the complexities of modern development, including the ability to install packages and run code directly, providing a practical workspace for testing the types of autonomous capabilities demonstrated by models like DaisyRose-27B.

Frequently asked questions

What training method did RadixArk use for DaisyRose-27B?

RadixArk used reinforcement learning exclusively for the post-training phase of DaisyRose-27B. Instead of relying on standard supervised fine-tuning, they applied this method to the Qwen3.8-27B base model to optimize its performance for agentic coding tasks.

What is the Terminal Bench 4.1 score for DaisyRose-27B?

DaisyRose-27B achieved a pass@1 score of 12.0 on the Terminal Bench 4.1. This represents more than a two-fold increase in performance compared to previous iterations, highlighting the model's effectiveness in command-line environments.

What hardware is required to run DaisyRose-27B?

The model is efficient enough to run on a single GPU. This allows developers to utilize the 27B parameter model without needing a multi-GPU setup, making it more accessible for individual researchers and smaller development teams.

Source: MarkTechPost


Source: MarkTechPost — RadixArk Ships DaisyRose-27B, a 27B Agentic Coding Model Trained With Reinforcement Learning Only