Zhi Spectrum (02513) announced interim results, revenue of 954 million yuan, up 399.7% year-on-year

According to Zhitong Finance App News, Zhitong Financial App (02513) announced the interim results for the six months ended June 30, 2026. The group obtained revenue of 954 million yuan, an increase of 399.7%; gross profit of 252 million yuan, an increase of 163.7%; R&D expenses of 2,131 million yuan, an increase of 33.64%; an adjusted net loss of 1,964 million yuan, an increase of 12.1%; and a loss per share.

In the first half of 2026, the company verified the deliverability of large models in terms of programming (coding). From early simple programming recommendations to complete code file generation to the implementation of an entire engineering project (Agentic Engineering), the company completely completed this curve: the model is no longer just a tool, but also has the ability to complete one thing independently. The revenue structure was then replaced. The share of revenue from open platforms and API services billed per call rose from 26.3% in the full year of last year to 86.5%, and revenue from localized model deployment fell from 73.7% to 13.5%. The main revenue structure will be replaced within the first half of 2026. In fact, the essence is something that will inevitably happen after the model's ability changes. Cybersecurity Cowork (Cybersecurity Cowork) is an extension of the model's same competency curve. Software bug mining does not rely on memory, but rather on the model's depth of reasoning and long-term autonomy. This is the direction in which programming ability naturally extends after crossing the threshold of an engineering project. During the reporting period, our model achieved 84.5% results on CyberGym (surpassing FABLE5 and GPT-5.6SOL) and completed 130 tasks on ExploitGym (second only to FABLE5 and GPT-5.6SOL).

Underpinning these two curves is infrastructure. Since the beginning of 2026, the company has fully implemented the all-in-infra strategy, invested in both training and inference, and incorporated domestic chips into the main inference computing power. It has achieved large-scale low-cost inference with 100,000-class domestic chips, and the unit token inference cost has decreased by 80% compared to the beginning of the year. Under conditions where the global supply of computing power continues to be tight, this path has determined that the company's cost curve is in its own hands. By the end of the reporting period, the total number of MaaS platform enterprises and developers using the GLM model exceeded 5.8 million. The company's total revenue was about 954 million yuan (approximately US$142 million), an increase of nearly 400% over the previous year. Among them, revenue from the open platform and API business was about 825 million yuan (approximately US$123 million), an increase of 2,735.7% over the previous year.

In August 2026, the company released GLM-5.3 and GLM-5.3-Flash (Ox-Alpha), and simultaneously opened GLM Coding Plan subscriptions to all users, covering Lite, Pro, Max and Team versions. This is one of the first large-scale model services in China to deliver the ability to code smart devices on a subscription basis. As of the date of this announcement, the GLM model has more than 7.4 million MaaS platform enterprises and developers.