In the second half of August, a model appeared on the OpenCode platform that initially no one could attribute to any specific laboratory. It was called Ox Alpha, was available for free, wrote above-average programming code, and offered no clues about its origin. After several days of speculation, the Chinese company Z.ai came forward and explained that it was its new GLM-5.3-Flash model, which it had been testing anonymously on the OpenCode and OpenRouter platforms.
The model appeared without documentation
OpenCode offered Ox Alpha as unidentified access to a model. It had no model card, parameter count, or name of the company that developed it. At the same time, the platform warned that Ox Alpha was merely a code name and that the underlying system could change at any time without warning, so its characteristics could not be relied upon in the long term.
Nevertheless, the published specifications attracted attention. The model supports a context window of one million tokens, can work with images and video, and OpenCode added an assurance that it does not retain any data from its operation. During the free period, the developers advertised a capacity of one hundred trillion tokens per day.
It excelled at visual tasks in tests
In an independent benchmark known as King Bench, Ox Alpha scored 70 out of 80 points and placed second overall. It received full marks for rendering a contact lens in the Three.js library and generating an SVG illustration of a panda, while earning eight out of ten points for an elevator simulation and a bow-and-arrow game.
It was this distribution of points that developers noticed most. Most models can write functional logic, but the result often looks like a typical website generated by artificial intelligence. In practical tests, Ox Alpha produced pages with clean typography, restrained use of color, and animations that did not feel like unnecessary decoration.
The model was released by Z.ai
Some members of the community guessed that it came from one of the Chinese laboratories, while others believed Microsoft was behind the model. Speculation also pointed to a new generation of GLM models, mainly because of its strong programming performance and pricing. The mystery also spread to Chinese social media, where users began calling the model Niu Lai after a low-budget animated film about a calf that had become an internet sensation at the time. On Wednesday, Z.ai ended the speculation and confirmed that Ox Alpha was actually GLM-5.3-Flash. The model weights are publicly available on the Hugging Face platform.
Z.ai added that all traffic during the Ox Alpha model testing was handled by Chinese artificial intelligence chips. Although it acknowledged their weaknesses in memory capacity and bandwidth, its engineers were able to compensate for them through software optimization and efficient workload distribution across individual machines. Washington has long restricted China’s access to the most powerful hardware for developing and operating artificial intelligence. This was not the first such experiment. The company had previously tested the GLM-5 model anonymously under the name Pony Alpha. According to the company, the main objective with GLM-5.3-Flash was to collect feedback from real users before its official market launch.
The company grew out of a Beijing university
Z.ai, which is also known in China as Zhipu AI, was founded in 2019 on the basis of technology developed at the prestigious Tsinghua University in Beijing. It used this technology to build the GLM family of large language models, an abbreviation for General Language Model, and launched its own ChatGLM chatbot in 2023.
It sells its services in two ways. It either provides them to customers through the cloud or deploys customized models directly on their own infrastructure. Its offering includes application programming interfaces, enterprise agents, general-purpose models, and technical support. In the domestic market, the company competes with DeepSeek, Moonshot AI, which is behind the Kimi chatbot, and Alibaba’s Qwen team. These projects compete on both performance and price, while all of them are capable of coping with limited computing power.
The company went public in January of this year, still under the name Knowledge Atlas Technology, which it changed to Z.ai in July. It offered its shares at 116 Hong Kong dollars, and they are currently trading at approximately 1,100 Hong Kong dollars.
Last year, Z.ai generated revenue of more than 720 million yuan, more than double the figure from 2024. At the same time, it invested around 3.2 billion yuan in research and development and posted a net loss of approximately 4.7 billion yuan.
Sources: businessinsider.com and mindstudio.ai
