GLM-5: from Vibe Coding to Agentic Engineering
GLM-5 Team, Aohan Zeng, Xin Lv, Zhenyu Hou, Zhengxiao Du, Qinkai Zheng, Bin Chen, Da Yin, Chendi Ge, Chengxing Xie, Cunxiang Wang, Gengzheng Pan, Hao Zeng, Haoke Zhang, Haoran Wang, Huilong Chen, Jiajie Zhang, Jian Jiao, Jiaqi Guo, Jingsen Wang, Jingzhao Du, Jinzhu Wu
2026-02-18
Summary
This paper introduces GLM-5, a new and improved AI model that's designed to be more than just a text generator – it's built to act as an intelligent agent capable of complex tasks, especially in coding.
What's the problem?
Previous AI models, while good at generating text or code, were often expensive to train and run, and struggled with tasks requiring long-term planning or understanding complex instructions. They also weren't very good at learning from their mistakes and improving over time, especially in real-world scenarios like software development.
What's the solution?
The creators of GLM-5 tackled these issues by using a new technique called DSA to make the model more efficient, reducing the computational resources needed for both training and using it. They also developed a new system for 'teaching' the model through reinforcement learning, allowing it to learn from experience without constantly needing to be retrained from scratch. This new learning system lets the model handle more complicated, long-term tasks more effectively.
Why it matters?
GLM-5 represents a significant step forward in AI because it's not just about generating outputs, but about creating an AI that can actually *do* things, particularly in the field of software engineering. It performs better than previous models on standard tests and, crucially, shows a real ability to handle complete software projects, making it a potentially valuable tool for developers.
Abstract
We present GLM-5, a next-generation foundation model designed to transition the paradigm of vibe coding to agentic engineering. Building upon the agentic, reasoning, and coding (ARC) capabilities of its predecessor, GLM-5 adopts DSA to significantly reduce training and inference costs while maintaining long-context fidelity. To advance model alignment and autonomy, we implement a new asynchronous reinforcement learning infrastructure that drastically improves post-training efficiency by decoupling generation from training. Furthermore, we propose novel asynchronous agent RL algorithms that further improve RL quality, enabling the model to learn from complex, long-horizon interactions more effectively. Through these innovations, GLM-5 achieves state-of-the-art performance on major open benchmarks. Most critically, GLM-5 demonstrates unprecedented capability in real-world coding tasks, surpassing previous baselines in handling end-to-end software engineering challenges. Code, models, and more information are available at https://github.com/zai-org/GLM-5.