Zhipu (Z.ai) releases GLM-4.5 series

Zhipu AI, trading as Z.ai, released the GLM-4.5 model family: a flagship 355-billion-parameter mixture-of-experts model and a smaller 106-billion-parameter GLM-4.5-Air variant, both under an open, auditable licence permitting on-premise deployment and fine-tuning. The company described GLM-4.5 as its first foundation model built with agent capability native to the architecture, designed to plan multi-step tasks and manage workflows end-to-end rather than have agentic behaviour layered on top of a general chat model.

Z.ai reported that GLM-4.5 placed third overall, and first among domestic and open-source models, on an aggregate of twelve benchmarks spanning reasoning, coding and agentic tasks — a comparison drawn from the company’s own benchmark selection rather than an independent evaluation. It priced API access at $0.11 per million input tokens and $0.28 per million output tokens, positioning the release on cost as much as capability, with the company framing its argument as showing that frontier-level performance could be delivered openly and affordably rather than only through closed, expensively priced models.

The release came six days after Alibaba’s Qwen3-Coder and roughly six months after DeepSeek’s R1 had first demonstrated that Chinese labs could ship open-weight models competitive with closed Western frontier systems at a fraction of the reported cost. GLM-4.5 extended that pattern into agentic use specifically, adding to a summer in which several Chinese labs released open models built around the same premise: that agentic capability, not just raw benchmark scores, was becoming the dimension on which open and closed model families competed most directly.