A new AI model has been introduced in China. Called GLM-4.5, it is an open-source model opposite to the most of the US-based AI system that are closed, and some benchmarks have put it ahead of Deepsek R1 and Chatgipt. GLM 4.5 has been developed by the Startup Z.Ai (east known as Zhipu). According to the company, this model is specifically designed for intelligent agent tasks and uses it described as “agentic” AI architecture. This means that the AI model can autonomally take tasks and handle logic, coding and other applications more effectively.

According to the company, GLM-4.5 is a large language model featuring 355 billion parameters, with a customized version called GLM-4.5 Air with 106 billion parameters, which makes it lighter and faster. The model supports a reference window of 128,000 tokens, allowing it to process long interactions or documents without losing attention. It is also said that indigenous function involves calling, which enables spontaneous integration with external software and workflow. The company highlights the fact that these abilities suit GLM-4.5 for many types of applications including advanced coding, physics simulation, game development and interactive mini-game.

The company has also revealed that GLM 4.5 has been issued under the Apache 2.0 Open-SOS license. This means that it is free to use and developers can download and deploy it independently, CNBC said.

Z.AI claims that GLM4.5 is globally ranked third and the 12 major AI ranks first between Chinese and Open-SOS model in benchmarks. According to the company, the model scored 98.2 percent on the Math500 Reasoning Test and 91 percent on Aime24 Challenge. It also gave 64.2 percent accuracy at a benchmark used for SWE-Bench verified, software engineering works, and achieved 90.6 percent tool-validity success rate by excluding the leading rivals.

Navidia operated by chip

One of the biggest benefits of GLM4.5 is the cost. Z.AI CEO Zhangpeng told CNBC that the model could run on only eight NVidia H20 chips. These chips are specially designed for the Chinese market under American export control. It is about half of the hardware required by the comparable model of deepsake. Jhangpeng revealed that the company does not currently need to buy additional chips, showing that the model already has enough computing capacity.

Cheaper

The company revealed that it had aggressively cut tokens pricing. Z.Ai will charge a fee of $ 0.11 per million input tokens and $ 0.28 per million output tokens compared to $ 0.14. This is less than $ 2.19 charged by Deepsek. In particular, tokens are the standard unit of data measurement for AI models.

For reference, Deepsek, the advanced LLM launched earlier this year, has been developed by the Chinese Startup High Flire AI and is known to optimize the chat of OpenaiI in the understanding and logic of natural language, while allegedly requires training cost under $ 6 million. Although GLM-4.5 is about half of the deepsac, but it is deferred to use agent AI design to maintain high accuracy and flexibility in completing tasks with low computational resources.

Interestingly, the arrival of GLM-4.5 occurs at a time when the development of China’s AI model increases rapidly. By July, Chinese companies had released 1,509 big language models compared to any other country, according to the state -owned Xinhua news agency.