Beijing-based Z.ai, recently added to the U.S. Entity List and backed by $1.5 billion in fresh funding, has released two open-source AI models that outperformed most Western counterparts in global benchmarks, challenging U.S. sanctions as it prepares for a Hong Kong listing.
The startup, formerly known as Zhipu AI, is listed on Washington’s Entity List, a trade restriction tool that identifies foreign individuals, companies, and other organizations the government believes are involved in activities contrary to U.S. national security or foreign policy interests.
Rubbing salt on the wounds, Z.ai dunked on San Francisco-based OpenAI in a blog post earlier on Monday, arguing that the massively popular AI Platform is something of a mess, with a confusing array of special-use products.
By comparison, Z.ai’s new open-source model GLM-4.5 merges reasoning and non-reasoning skills into one fully capable and versatile architecture.
“In the past five years, OpenAI’s GPT-3 learns common-sense knowledge, and o1 uses reinforcement learning to think before responding, significantly improving reasoning skills in coding, data analysis, and complex math,” the post said.
“However, the resultant models are still not really general: some of them are good at coding, some good at math, and some good at reasoning, but none of them could achieve the best performance across all the different tasks,” it said in its post. “GLM-4.5 makes efforts toward the goal of unifying all the different capabilities.”
OpenAI has stated that its next model, GPT-5, expected by the end of summer, will be its first fully unified tool.
In the meantime, GLM-4.5 and its smaller sibling GLM-4.5-Air scored an average of 63.2 and 59.8 across 12 industry tests, ranking third globally, behind OpenAI o3 and Grok-4.
The models employ a Mixture of Experts architecture with a total of 355 billion parameters for the flagship version. However, only 32 billion are active at any given time, making it more efficient than brute-force approaches.
Image: z.AI
GLM-4.5 achieved a 90.6% tool-calling success rate, outperforming Claude 3.5 Sonnet’s 89.5% and even Kimi K2’s 86.2%.
On web browsing tasks, it correctly answered 26.4% of complex questions compared to Claude 4 Opus’s 18.8%. The model solved 98.2% of problems in the MATH 500 benchmark, matching Claude 4 Opus, all while being open source.
The U.S. Department of Commerce added Zhipu to its Entity List in January, accusing the company of supporting Beijing’s military advances.
Companies on the list cannot purchase American technology without special government approval. Zhipu said the designation “will not have a substantial impact” on its operations.
Which is undoubtedly true: Thus far, State-backed investors have poured money into the company since the Entity List announcement.
Huafa Group, backed by the Zhuhai government, invested 500 million yuan ($69 million) in March. Chengdu’s state fund followed with 300 million yuan, while Shanghai’s Pudong Venture Capital Group and Zhangjiang Group jointly added 1 billion yuan in July.
The startup plans to raise another $300 million through a Hong Kong initial public offering, according to people familiar with the matter.
The company initially filed for a mainland China listing but shifted strategies as geopolitical tensions escalated. Its valuation doubled to 40 billion yuan ($5.6 billion) between September 2024 and July 2025.
Z.ai prices API calls at $0.11 per million input tokens and $0.28 per million output tokens—roughly one-fifth the cost of comparable Western models. The company released the weights under an MIT license on Hugging Face and GitHub, allowing anyone to download and modify the code.
The models feature what Z.ai calls “hybrid reasoning” with two modes.
A thinking mode handles complex reasoning and planning tasks, while a non-thinking mode provides instant responses.
The company shared samples of how people can build entire websites with natural language commands or generate interactive visualizations ranging from Pokémon databases to complex presentations.
Chinese companies are accelerating their investment in AI, releasing 1,509 large language models as of July 2025, which represents 40% of the 3,755 models released globally and offers some of the best open-source models in the industry.
Just recently, Kimi K2 impressed. Additionally, today, Alibaba released Wan 2.2, which is 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 beating some of the biggest names in the generative video scene.
Early investors of Z.ai include Alibaba Group, Tencent Holdings, and HongShan Capital Group. The company spun out from Tsinghua University in 2019 and represents one of China’s “AI Tigers,” startups that the government considers crucial for achieving technological independence amid heavy political constraints imposed by the U.S.
To boost its own ecosystem of AI technology and apps, Z.ai launched a 1.5 billion yuan ($211 million) fund called the “Z Fund” to invest in startups that build on its technology.
The strategy mirrors OpenAI’s approach of creating a developer ecosystem around proprietary models while maintaining more open licensing terms.
Users can try an online version of the chatbot by accessing this link: Chat with Z.ai. The site is hosted in Singapore, so you can breathe a little easier.
Read the full article here
