header-langage
简体中文
繁體中文
English
Tiếng Việt
한국어
日本語
ภาษาไทย
Türkçe
Scan to Download the APP
Partnered With:

SmartLedger's Coding Turnaround: DeepSeek Forces It to Switch Direction, Users Chart the Course

Insightful Beating AI News Flash: 36Kr reviewed the behind-the-scenes process of ZhiPu's transformation through Coding:

After the Lunar New Year in 2025, DeepSeek R1 suddenly gained popularity, directly impacting ZhiPu's core custom enterprise business. An employee estimated that nearly 30% of the clients switched to DeepSeek. In order to retain these clients, ZhiPu once offered significant discounts.

In May, ZhiPu held a crucial strategic meeting. The company abandoned the separate routes of text, multimodal, and programming to develop vertical models, deciding to integrate Reasoning, Coding, and Agentic capabilities into one large model. This temporary shift also delayed the release of GLM-4.5 from April to July.

Interestingly, ZhiPu initially did not prioritize Coding. The three capabilities were initially equal, and it was actually the users who pushed the model towards Coding. Throughout the development process, the most concentrated customer feedback was always about "improving development efficiency." After the launch of GLM-4.5, an increasing number of enterprises requested its integration into the programmer workflow, and developers started using it as a substitute for Claude. Coding quickly became the most well-received capability of GLM-4.5.

Subsequently, ZhiPu began to streamline its focus. The Coding and text directions currently have around a hundred people, while the multimodal training team has been reduced to 10 people. Training became more focused on data, post-training, and real customer feedback. When faced with uncertain technical directions, several research teams were tasked with running A/B tests separately, and the one with the best performance was chosen. Before model release, both general capabilities were tested and the delivery team conducted another round of testing in real-world scenarios.

After GLM-4.5, ZhiPu introduced the Coding Plan and Enterprise Edition, gradually making Coding one of the most important capabilities in the GLM series. By the time GLM-5.2 was released, the model's reputation and API revenue soared. ZhiPu's Annual Recurring Revenue (ARR) jumped from $500 million to $600 million in May of this year, and suddenly rose to $1 billion in July.

ZhiPu did not initially focus on Coding. DeepSeek first disrupted its comfort zone, and then customers chose a new direction for it.

Correction/Report
Submit
Add Library
Visible to myself only
Public
Save
Choose Library
Add Library
Cancel
Finish