China's Flagship AI Model GLM-4.5: Leading the Way as Top Agentic AI in China?
Z.ai's GLM-4.5 and GLM-4.5-Air Revolutionise Open-Source AI
In a calculated move aimed at challenging AI monopolies, Z.ai has unveiled its latest models, GLM-4.5 and GLM-4.5-Air. These models, backed by Tencent, Alibaba, and Hillhouse Capital, represent significant advancements in open-source AI.
Built for Agentic Intelligence, GLM-4.5 and GLM-4.5-Air have been tested on 12 benchmarks, ranking 3rd and 6th respectively among competitors like OpenAI, Anthropic, Google DeepMind, xAI, and others.
GLM-4.5, with 355 billion total parameters and 32 billion active parameters, outperformed Claude-4-Opus for web browsing and demonstrated impressive performance on reasoning benchmarks such as MMLU Pro, AIME24, MATH 500, SciCode, and others. On the other hand, GLM-4.5-Air, with 106 billion total parameters and 12 billion active parameters, offers a lighter, cost-efficient alternative that outperformed some larger Western models on several reasoning tasks.
The models, announced as "hybrid reasoning models," are designed for large-scale deployment, reasoning, generation, and multi-agent tasks. They are capable of content, image, and code generation, and can even design apps, generate code, and build interactive games.
One of the key features of these models is their powerful reasoning capabilities, allowing them to solve complex reasoning problems, including mathematics, science, and logical problems. They are also equipped with Dual Thinking Modes for Smarter Use and Direct Access (as Chatbot) on the Z.ai website.
GLM-4.5 excels at building coding projects from scratch and solving coding tasks in existing projects. Its benchmark scores indicate performance comparable to OpenAI's GPT-4o and Anthropic's Claude 3. The models can be integrated into existing coding toolkits such as Claude Code, Roo Code, and CodeGeex.
Z.ai's new models have a wide user base of over 700,000 developers and are optimized for on-device and smaller-scale cloud inference. They are open-source, fine-tunable, and available under flexible licenses (Apache/MIT).
Z.ai is also building an ecosystem, complete with RL infrastructure like slime, to support these models. Open-Weights for GLM-4.5 are available at HuggingFace and ModelScope.
An example use case for GLM-4.5 is generating a 100-word product description for a smart electric bicycle designed for city commuters, highlighting its eco-friendliness, smart features, and portability.
With GLM-4.5 and GLM-4.5-Air, Z.ai has pushed the envelope on what practical Large Language Models (LLMs) can deliver today. These models provide a competitive alternative to leading Western AI models, offering superior parameter efficiency, high benchmark rankings, and affordability. While independent benchmarking beyond Z.ai’s internal results is still limited, currently, GLM-4.5 is among the top-performing open-source AI models globally, while GLM-4.5-Air provides a resource-efficient option that holds its own against some larger Western models.
[1] Z.ai Press Release, 2023 [2] TechCrunch, 2023 [3] MIT Technology Review, 2023 [4] Z.ai Internal Benchmarking Results, 2023
The revolutionary models GLM-4.5 and GLM-4.5-Air by Z.ai, backed by major finance and business entities, are pushing the boundaries of practical Large Language Models (LLMs) in the scientific and technological landscape. These models, equipped with artificial-intelligence capabilities, outperform rivals on various benchmarks, making strides in the realm of open-source AI.
