GLM-4.5
Zhipu AI (Z.ai) · 2025-07-28 · 358.3B parameters
GLM-4.5 is Z.ai's July 2025 MIT-licensed open-weight Mixture-of-Experts language model for agentic, reasoning, and coding tasks, with chat-template-selectable thinking and non-thinking modes. Z.ai markets it as 355B total and 32B activated parameters, while the released checkpoint contains exactly 358,337,776,896 model-weight parameters; its 92-layer main stack uses GQA with QK-Norm, partial RoPE, and a configured 131,072-token maximum, plus one MoE MTP layer for speculative decoding. The report describes a roughly 23T-token pre-training and mid-training curriculum followed by supervised fine-tuning and reinforcement learning, so exact cumulative training exposure is not disclosed.