36Kr recapped the story behind Zhipu's coding-powered comeback. It didn't start as a coding triumph. When DeepSeek R1 blew up over the 2025 Lunar New Year, the enterprise customization business Zhipu relied on most took a direct hit. One employee estimated that nearly 30% of customers switched to DeepSeek. To hold onto the rest, Zhipu handed out steep discounts.
Then, in May, the company held a decisive strategy meeting. It abandoned the plan to build separate vertical models for text, multimodal, and coding, and instead decided to cram Reasoning, Coding, and Agentic capabilities into one large model. That sudden re-route is why GLM-4.5, originally slated for April, didn't launch until July.
Funny thing: Zhipu wasn't sold on coding at first, either. The three capability tracks had no obvious pecking order. Users made the call. Throughout development, the most common client request was "improve R&D efficiency." After GLM-4.5 shipped, more and more companies asked to integrate it into their programmers' workflows, and developers started using it as a cheaper Claude alternative. Coding quickly became GLM-4.5's best-received feature.
Zhipu then tightened its focus. Coding and text teams now have around 100 people; the multimodal training team is down to 10. Training now leans harder on data, post-training, and real customer feedback. When a technical direction is murky, several research groups run A/B tests, and whatever works best wins. Before a release, Zhipu benchmarks general capability, then has its delivery team re-test the model against real scenarios.
After GLM-4.5, Zhipu launched a Coding Plan and an enterprise edition, turning coding into one of the GLM family's most important capabilities. By GLM-5.2, both model reputation and API revenue were taking off. Zhipu's annual recurring revenue jumped from $500 million to $600 million in May to $1 billion in July.
Zhipu didn't bet on coding from the start. DeepSeek first knocked out its comfort zone — then users drew the new map.