On Monday, the Chinese AI lab Z.ai unveiled two new open-source general language models (GLM), named GLM-4.5 and GLM-4.5-Air. The company identifies these models as its latest flagship offerings, designed to enhance the capabilities of artificial intelligence. Both models integrate hybrid reasoning approaches, operating with a thinking mode for complex tasks and a non-thinking mode for quick responses. The company also highlights that its new models possess agentic attributes and claims they surpass all other available open-source models globally.
Z.ai Unveils Open-Source GLM-4.5 AI Models
In a recent blog post, Z.ai elaborated on the goal of its GLM models, which is to develop a versatile language model capable of effectively tackling a range of tasks. The firm asserts that, despite significant progress in generative AI by leading competitors such as Google, OpenAI, and Anthropic, their models do not exhibit true generality, excelling in some areas while underperforming in others. “GLM-4.5 aims to unify diverse capabilities,” the company stated.
The GLM-4.5 model boasts 355 billion total parameters, of which 32 billion are active, while the Air variant features 106 billion parameters, with 12 billion active. Both models are designed to integrate reasoning, coding, and agentic capabilities within a single framework, featuring an expansive context window of 128,000 tokens and an inbuilt function calling feature.
To enhance computational efficiency during training and inference, Z.ai adopted a mixture-of-experts (MoE) architecture for the model’s framework. Rather than expanding the width of the MoE layers, as seen in DeepSeek-V3, the GLM-4.5 series opts for a narrower configuration while increasing the number of layers. The rationale is that deeper architectures tend to exhibit enhanced reasoning skills.
The company also disclosed innovative techniques employed during the pre-training and post-training phases in its blog post, aiming to provide transparency for the developer community regarding the models’ construction.
Performance of GLM-4.5 series AI models
Photo Credit: Z.ai
Z.ai reported that it evaluated the GLM-4.5 model across 12 benchmarks involving agentic tasks, reasoning, and coding, claiming that the model’s performance was compared to various LLMs from competitors like OpenAI, Anthropic, Google, xAI, and Alibaba. This internal assessment placed GLM-4.5 in the third position overall, trailing OpenAI’s o3 and xAI’s Grok 4.
For those interested in exploring these models, Z.ai has made the open weights available on its GitHub page and on Hugging Face’s collection. Additionally, access to these LLMs is possible through the company’s website or application programming interface (API).
