GLM-4.5 Launches with Strong Reasoning, Coding, and Agentic Capabilities

Zhipu AI has released two new AI models, GLM-4.5 and GLM-4.5-Air, designed to efficiently perform reasoning, coding, and agent tasks. The models utilize a dual-mode system that enhances accuracy and response speed. GLM-4.5 has 355 billion parameters, while GLM-4.5-Air has 106 billion. They employ a Mixture-of-Experts architecture and a "thinking" mode for complex tasks, alongside a "non-thinking" mode for faster outputs. Training utilized a 22 trillion-token corpus, excelling in coding benchmarks with significant success rates, surpassing many peer models. Both models ranked high in various task benchmarks, evidencing their capabilities.

GLM-4.5 achieved 64.2% on SWE-bench Verified and 37.5% on TerminalBench, outperforming Claude 4 Opus, GPT-4.1, and Gemini 2.5 Pro on several metrics.

GLM-4.5 incorporates QK-Norm, Grouped Query Attention, Multi-Token Prediction, and the Muon optimizer for faster convergence and improved reasoning performance.

Read at InfoQ