‘Mysterious AI’ Ox Alpha uses entirely Chinese chips

On August 27, startup Z.AI, also known as Zhipu AI in China, officially launched the open source model version GLM-5.3-Flash, confirming that this is Ox Alpha. Previously, Ox Alpha was tested as a “secret model” from an anonymous third party on the OpenRouter platform, OpenCode to collect user feedback before release and quickly caused a stir globally.

At the announcement ceremony, Z.AI emphasized that GLM-5.3-Flash “fully operates” on a server cluster containing 100,000 domestic AI chips. The company did not specify the supplier, but according to China Dailythey currently cooperate with many semiconductor manufacturers such as Huawei, Cambricon Technologies and Moore Threads.

Also on August 27, Cambricon confirmed achieving compatibility “from day one” in serving GLM-5.3-Flash. Moore Threads also said it has provided “from the ground up” support for the new AI model.

SCMP assessment, the use of domestic chips “marks an important test of China’s ability to handle large-scale inference workloads on home-grown hardware”, as the country seeks to reduce its dependence on advanced processors from Nvidia.

According to Z.AI, with a total of 320 billion parameters, GLM-5.3-Flash only activates 18 billion parameters per request to minimize the computational burden. In order to overcome the low memory capacity and bandwidth that are the weaknesses of Chinese AI chips, the company builds a specialized inference engine, dividing the processing stage into independently managed computing groups. This architectural adjustment triples the service performance compared to normal, bringing hardware efficiency and cost per token to the same level as mainstream Nvidia accelerators. However, these claims have not been independently verified.

A token is a basic unit of input data consisting of a word, part of a word, character, or punctuation that an AI model processes to produce an output. According to Z.AI, after deployment on OpenRouter and OpenCode, GLM-5.3-Flash processed 62,000 billion tokens before officially launching, also the first model in the GLM-5 series capable of processing image information along with text naturally.

Independent records from OpenRouter show that the system processed more than 11,000 billion tokens in the first three days, equivalent to nearly 31% of the total weekly transaction volume. Performance evaluation company Artificial Analysis gave the model a score of 57, ranking it 10th globally and third among open-weighted models, behind only Moonshot AI’s Kimi K3 and Alibaba’s Qwen3.8 2.4T A95B, both from China.

Z.AI currently offers GLM-5.3-Flash at a competitive price to attract domestic and international developers. The cost to run 1 million tokens is $0.15 in and $0.50 out, which is 1/10 of the standard GLM-5.3, even 1/20 in the promotion, and about 1/40 of Anthropic’s Opus 4.8 “at the same level of intelligence”. TechCrunch Ox Alpha’s appearance increases the possibility of cheap, effective Chinese models taking significant market share from expensive American advanced model development companies such as OpenAI and Anthropic.

Theo CNBCZ.AI shares increased more than 12% after the trading session on August 27 on the Hong Kong stock exchange.

Despite recording high traffic during the free trial period, initial feedback from developers was mixed. Some praised the model’s ability to debug complex code, solving 28% of 175 problems on the LiveCodeBench programming competency assessment tool. Patrick Collison, CEO of technology company Stripe, tried the model and commented “very impressive”.

However, many programmers say that AI still has “illusion” phenomena, abandoned tasks or slow code creation. Artificial Analysis also rates data speeds as “slower than industry average”.

Z.AI (Zhipu AI) was founded in 2019 in Beijing, developed from a research team at Tsinghua University. The company mainly focuses on developing large language models and AI products of the GLM (General Language Model) series, and is one of the Chinese AI startups receiving investment from domestic technology funds and enterprises.