For several days, a prominent AI model garnered attention without a clear origin. Named Ox Alpha, this innovative tool was available for free, showcased impressive coding capabilities, and left developers speculating about the major AI lab responsible for its creation.
The veil of mystery was lifted when Z.ai, a Beijing-based AI company, stepped forward this week to announce that Ox Alpha was merely a pseudonym for its latest open-weight model, GLM-5.3-Flash. This model was tested anonymously on platforms like OpenRouter and OpenCode, propelling Z.ai into the limelight amid China's escalating AI competition.
Z.ai is now in competition with other notable players like DeepSeek, Moonshot AI (the creator of Kimi), and Alibaba's Qwen team. These companies are striving to advance not only the performance and cost-efficiency of their technologies but also to navigate the constraints of available computing resources.
Founded in 2019, Z.ai, also known as Zhipu AI in China, emerged from technology cultivated at Tsinghua University in Beijing. The company developed a series of large language models called GLM, launching its chatbot, ChatGLM, in 2023. In January, Z.ai made its public debut on the Hong Kong stock exchange under the corporate name Knowledge Atlas Technology; however, it rebranded to Z.ai in July. The company's shares, which were initially priced at 116.20 Hong Kong dollars when trading began on January 8, have surged to approximately HK$1,100, reflecting a ninefold increase since its IPO.
In the previous year, Z.ai reported revenue of 724.3 million Chinese yuan (around $108 million), more than double its expected figure for 2024. However, it incurred approximately 4.7 billion yuan in net losses, spending about 3.2 billion yuan on research and development. The company is set to release its financial results for the first half of the year on Monday.
Z.ai's operations are divided primarily into two segments: cloud-based services and tailored models that can be deployed on clients' own infrastructure. Its offerings encompass APIs, enterprise AI agents, general-purpose models, and technical assistance.
Ox Alpha's true identity as GLM-5.3-Flash sparked intrigue among developers, some of whom speculated whether a Chinese research lab or even Microsoft was behind the model. This speculation made its way into online discussions in China, where it was humorously dubbed the "Niu Lai Model" after a low-budget animated film featuring a calf that became an online meme.
Clarifying the situation on Wednesday, Z.ai confirmed that Ox Alpha was indeed its GLM-5.3-Flash model, which is multimodal and can process text, images, and video. The weights for this model are accessible on Hugging Face, enabling developers to download and utilize it.
Notably, Z.ai disclosed that the computing resources used during the Ox Alpha testing were entirely based on Chinese AI chips. This announcement comes in light of U.S. restrictions limiting China's access to some advanced chips essential for AI development. Although these chips have certain constraints in terms of memory capacity and bandwidth, Z.ai's engineers managed to optimize their software and efficiently allocated workloads across the hardware.
Z.ai has a history of testing models anonymously; prior to Ox Alpha, they experimented with the GLM-5 model under the name Pony Alpha. The intention behind the anonymous testing of GLM-5.3-Flash as Ox Alpha was to collect valuable user feedback prior to its official release.
