Chinese artificial intelligence (AI) lab Z.ai announced the release of two new open-source general language models (GLM) on Monday. Dubbed GLM-4.5 and GLM-4.5-Air, the AI firm calls them its latest flagship models. Both of them are hybrid reasoning models that offer a thinking mode for complex reasoning and tool use, and a non-thinking mode for instant responses. Additionally, the company says these models support agentic capabilities. Notably, the AI firm claims that its latest models outperform all other open-source models worldwide.
Z.ai Introduces Open-Source GLM-4.5 AI Models
In a blog post, the Chinese AI firm announced the release of these models. The idea behind GLM models is to create a large language model (LLM) that is truly generalistic and can equally perform different types of tasks. The company argues that despite several advancements in generative AI, the models from the likes of Google, OpenAI, and Anthropic are not general enough, as they display strong performance in some areas while lagging in others. “GLM-4.5 makes efforts toward the goal of unifying all the different capabilities,” the company said.
The GLM-4.5 AI model features a total of 355 billion parameters with 32 billion active parameters. On the other hand, the Air variant gets 106 billion total parameters, with 12 billion parameters being active. Both models unify reasoning, coding, and agentic capabilities in a single architecture. It has a context window of 1,28,000 tokens and comes with native function calling capacity.
Coming to the model architecture, Z.ai opted for a mixture-of-experts (MoE) architecture to improve the compute efficiency of both training and inference. Instead of increasing the width (hidden dimensions and number of experts) of the MoE layers like DeepSeek-V3, the GLM-4.5 series reduces the width while increasing the height (number of layers). This was done as the company believed that deeper models displayed improved reasoning capability.
The Chinese AI firm also listed the novel techniques used for the pre-training and post-training process in the blog post to help the developer community understand how the models were built from scratch.
Performance of GLM-4.5 series AI models
Photo Credit: Z.ai
Z.ai claimed to have tested the GLM-4.5 model’s performance on 12 benchmarks across agentic, reasoning, and coding. It then claimed to compare the model’s overall scores against various LLMs from OpenAI, Anthropic, Google, xAI, Alibaba, and more. Based on this internal evaluation, the Chinese AI firm claims that GLM-4.5 ranked in the 3rd place, behind OpenAI’s o3 and xAI’s Grok 4.
Interested individuals can access the open weights of these models from Z.ai’s GitHub and Hugging Face listings. Alternatively, these LLMs can also be accessed via the company’s website and application programming interface (API).
