The World Artificial Intelligence Conference began in Shanghai on Saturday, July 26th, 2025.
nuphoto | nuphoto | Getty Images
Beijing – Chinese companies are creating artificial intelligence models that are increasingly cheaper to use, reflecting key aspects of Deepseek’s market-shaking breakthrough.
Startup Z.ai, formerly known as Zhipu, announced on Monday that it will use less of its new GLM-4.5 AI model than DeepSeek. In contrast to the logic underlying the existing AI models, Z.AI said the new GLM-4.5 is built on what is known as the “agent” AI. This means that the model automatically splits the task into subtasks to complete more accurately.
The new model is also open source. This means it’s free for developers to download and use.
At about half the size of Deepseek’s models, the GLM-4.5 only requires eight NVIDIA H20 chips to operate, Z.AI CEO Zhang Peng told CNBC on Monday.
This is a chip nvidia customized for China to comply with US export controls. The chip maker said this month that the US will allow them to resume sales in China after a three-month hiatus, but it is unclear when their shipments will begin.
Zhang said the company doesn’t need to buy more chips because it has enough computing capabilities for now, but refused to share how much Z.AI spent training the AI model. Details will be made public later, he said.
In January, Deepseek was rattling global investors with its apparent ability to ignore US chip restrictions and create AI models that not only rival US-based Openai’s ChatGPT, but also weaken it with training and operational costs.
Deepseek claimed that the training cost for the V3 model is less than $6 million, but some analysts say the figure is based on the company’s hardware spending exceeding $500 million over time.
Z.AI said the new GLM-4.5 model will charge 14 cents for a 14 cent input token for the DeepSeek R1. For DeepSeek, it’s $2.19 for an output token of 28 cents per million. Tokens are methods of measuring data from AI model processing.
Earlier this month, Alibaba-backed Moonshot released Kimi K2. This claimed to have defeated Openai’s ChatGpt and Claude of Mankind with a specific coding feature. Kimi K2 charges 15 cents per input token and $2.50 per output token, according to its website.
Weekly analysis and insights from Asia’s biggest economy in your inbox
Subscribe now
In late June, Openai warned Zhipu about the progress of China’s AI. The US has also added startups to its entities list that restrict American companies from doing business.
Z.AI was launched in 2019 and is reportedly planning an initial public offering for China’s Greatest.
According to Pitchbook, the startup has raised more than $1.5 billion from investors, including Alibaba, Tencent and Qiming Venture Partners. Aramco-backed Prosperity7 ventures and local government funding from cities in Hangzhou and Chengdu are also one of Z.AI’s supporters, the database shows.
Over the past few weeks, several other Chinese companies have unveiled new open source AI models. During the World AI Conference held in Shanghai, Tencent released the Hunyuanworld-1.0 model to generate 3D scenes for game development. Last week, Alibaba announced the QWEN3-Coder model for writing computer code.
