GLM-5 is a next-generation foundation model designed to shift software development from vibe coding toward agentic engineering. Built by Zhipu AI in collaboration with Tsinghua University, the model adopts Dense Sparse Anything (DSA) architecture to cut training and inference costs while preserving long-context fidelity. The design targets complex systems engineering and long-horizon agentic tasks, leveraging the agentic, reasoning, and coding foundations of its predecessor to handle end-to-end software engineering challenges that go beyond single-file generation.

Training leverages a new asynchronous reinforcement learning infrastructure that decouples generation from training, dramatically improving post-training efficiency. Novel asynchronous agent RL algorithms enable the model to learn from complex, multi-step interactions, driving state-of-the-art performance on major open benchmarks. On core agentic programming benchmarks like SWE-bench Verified and Terminal Bench 2.0, GLM-5 reaches open-source SOTA performance on par with leading proprietary models. Its ability to sustain optimization across hundreds of reasoning rounds and thousands of tool calls makes it especially effective for real-world coding tasks, autonomous debugging, and terminal-based automation workflows.