In short
- Z.AI’s GLM-4.5 and Aerial models jump western Ais and hit no. 3 in industrial tests worldwide.
- The startup, on the black list of Washington, raised $ 1.5 billion when Chinese state funds accumulated.
- New models are open-source, super efficient and cost a fifth of the Western rivals to use.
Z.AI, recently added to the American entity list and supported by $ 1.5 billion in new financing, has released two open-source AI models that performed better than most Western counterparts in global benchmarks, who challenge American sanctions while it is preparing on a list of Hong Kong.
The startup, Previously known as Zchipu AI, is stated on the Washington entity list, a tool for commercial restriction that identifies foreign persons, companies and other organizations The government is of the opinion that is involved in activities that are contrary to the American national security or foreign policy interests.
Wings salt on the wounds, Dunkte Z.AI on Monday on San Francisco in a blog post in a blog post, with the argument that the massively popular AI platform is something of a mess, with a confusing series of products for special use.
For comparison: Z.AI’s new open-source model GLM-4.5 combines reasoning and non-reasonable skills in one completely capable and versatile architecture.
“Over the past five years, the GPT-3 of OpenAI has learned common sense, and O1 uses reinforcement to think before they respond, considerably improving reasoning skills in coding, data analysis and complex mathematics,” the post said.
“However, the resulting models are still not really general: some are good in coding, some good in mathematics and some good in reasoning, but none of them could achieve the best performance in all different tasks,” said it in his post. “GLM-4.5 makes efforts in the direction of the goal of uniting all different options.”
OpenAi has stated that his next model, GPT-5, expects at the end of the summer, will be the first completely uniform tool.
In the meantime, GLM-4.5 and its smaller brother or sister GLM-4.5 ART scored an average of 63.2 and 59.8 in 12 industrial tests, which become third worldwide, behind OpenAI O3 and Grok-4.
The models use a combination of expert architecture with a total of 355 billion parameters for the flagship version. However, only 32 billion is active at a certain moment, making it more efficient than brutal-force approaches.
GLM-4.5 achieved a 90.6% Tools-return success rate, better than Claude 3.5 sonnets 89.5% and even Kimi K2’s 86.2%.
On Internet tasks, the 26.4% of the complex questions correctly answered 18.8% of Claude 4 Opus. The model solved 98.2% of the problems in the Math 500 benchmark, matching Claude 4 Opus, all while he was open source.
The US Department of Trade added Zchipu to his entity list in January and accused the company of supporting Beijing’s military progress.
Companies on the list cannot buy American technology without special government approval. Zchipu said that the designation “will not have a substantial impact” on its activities.
That is undoubtedly true: so far, the state has put money back in the company since the announcement of the entity list.
Huafa Group, supported by the Zhuhai government, invested 500 million Yuan ($ 69 million) in March. The Staatsfonds of Chengdu followed with 300 million Yuan, while the Pudong Venture Capital Group of Shanghai and Zhangjiang Group jointly added 1 billion Yuan in July.
The startup is planning to pick up another $ 300 million through an initial public offer from Hong Kong, according to people who are familiar with the issue.
The company initially presented a mainland list, but shifted strategies as geopolitical tensions escalated. The appreciation doubled between September 2024 and July 2025 to 40 billion Yuan ($ 5.6 billion).
Z.AI API calls at $ 0.11 per million input tokens and $ 0.28 per million output tokens series one fifth the costs of comparable Western models. The company has released the weights under a MIT license on hugging Face and Github, so that everyone can download and change the code.
The models contain what Z.AI call “hybrid reasoning” with two modes.
A thinking mode deals with complex reasoning and planning tasks, while a non-thinking mode offers immediate answers.
The company shared examples of how people can build entire websites with natural language assignments or generate interactive visualizations, ranging from Pokémon databases to complex presentations.
Chinese companies accelerate their investment in AI, so that 1,509 large language models are released from July 2025, which represents 40% of the 3,755 models that are released worldwide and offer some of the best open-source models in industry.
Kimi K2 was recently impressed. Moreover, Alibaba WAN 2.2 released today, the state-of-the-art (SOTA) Open-source generative video model.
Hidream-I1 is the best open-source image generator with hailuo, Seedream and Kling who beat some of the biggest names in the generative video scene.
Early investors from Z.AI are Alibaba Group, Tencent Holdings and Hongshan Capital Group. The company was based on Tsinghua University in 2019 and represents one of the “AI Tigers” of China, startups that the government considers crucial to achieve technological independence in the midst of heavy political restrictions imposed by the US
To stimulate his own ecosystem from AI Technology and Apps launched a fund of 1.5 billion Yuan ($ 211 million) called the “Z Fund” to invest in startups that build About his technology.
The strategy reflects OpenAI’s approach to create a developer ecosystem around its own models while retaining more open licensing conditions.
Users can try an online version of the chatbot by gaining access to this link: chat with z.ai. The site is hosted in Singapore, so you can breathe a little easier.
Generally intelligent Newsletter
A weekly AI trip told by Gen, a generative AI model.
