(Yicai) Aug. 27 -- Z.AI’s shares surged after the Chinese startup claimed responsibility for creating Ox Alpha, a powerful artificial intelligence model that dropped anonymously on application programming interface platforms OpenRouter and OpenCode a week ago and quickly set new AI usage records.

Z.AI, formerly known as Zhipu AI, [HKG: 2513] finished 12.6 percent higher at HKD1,160 (USD147.99) per share on the Hong Kong Stock Exchange today. The shares, which debuted on the bourse in January, have increased 10 times in value from their offering price of HKD116.20 (USD14.82).

Ox Alpha has been renamed GLM-5.3-Flash, making it the first native multimodal model in Z.AI’s GLM-5 series, the Beijing-based company announced in a press release late yesterday.

It underwent large-scale testing on OpenRouter and OpenCode before release, Z.AI said. Over the past week, the model set new records for call volume on both platforms, processing tens of trillions of tokens from developers around the world. All of the request traffic was supported by computing power using Chinese-made chips.

As of Aug. 25, GLM-5.3-Flash had processed about 23.2 trillion tokens within OpenRouter’s most recent seven-day statistical window, more than double that of its closest competitor DeepSeek-V4-Flash-0731, which had processed around 11.6 trillion in the period. However, the statistics measure only token usage and not user numbers, request counts, task success rates, or income.

For Z.AI, this is its first attempt at using a Chinese chip cluster under large-scale traffic conditions, the company said. Hardware efficiency and per-token cost have reached levels comparable to mainstream Nvidia graphics processing units at the cluster level, proving that China-made chips can fully support the inference demands of cutting-edge models efficiently and economically, it noted, without disclosing which specific chips are used.

In benchmark tests and real-world usage, GLM-5.3-Flash comprehensively outperforms GLM-5.2, which has twice its parameter count, while priced at only one-10th, as well as one-20th of GLM-5.3 within a limited discount period, and one-40th of Anthropic’s Opus 4.8, Z.AI pointed out. The architecture was designed specifically for extremely low cost, enabling stronger performance with fewer computational resources, it added.

OpenCode announced that Ox Alpha would be made available for free for one week with a “near-unlimited” quota when the model was released, adding that its provider had prepared service capacity of 100 trillion tokens per day. Developers flocked to test the model due to the free use while speculating which company possessed such massive inference resources.

In February, Z.AI used a similar approach to test GLM-5 after first anonymously releasing an early version of the model, called Pony Alpha, on OpenRouter.

Editor: Martin Kadiev