Z.ai has unveiled GLM-5.1, their next-generation flagship model for agentic engineering that delivers substantially stronger coding capabilities than its predecessor. The model achieves state-of-the-art performance on SWE-Bench Pro and leads benchmarks for repository generation and terminal tasks. Unlike previous models that plateau early, GLM-5.1 maintains effectiveness over extended task horizons.

Performance tests across three scenarios demonstrate GLM-5.1's long-horizon optimization capabilities. In vector database optimization, the model reached 21.5k QPS—roughly 6× the best single-session result—through 600+ iterations. For GPU kernel optimization, it achieved 3.6× speedup and continued improving well into extended runs. The model breaks complex problems down, runs experiments, and refines strategies through repeated iteration.

When tasked with building a Linux desktop as a web application, GLM-5.1 worked for 8 hours, continuously reviewing and improving its output. Early versions of GLM would have stopped after creating a basic skeleton, but GLM-5.1 identified missing features, refined styling, and fixed broken interactions autonomously. This sustained productivity over extended sessions represents a significant advance in AI agentic capabilities.