The term Vibe Coding (programming by feel, or agent-driven coding) is reshaping how software gets built and pushing valuations into the billions of dollars in Silicon Valley. Behind the hype, however, many of these breakthrough tools may be running on core AI models developed in China, a detail their founders do not discuss publicly. The tension between innovation, national tech pride, and transparency is now at the center of one of the fastest-growing trends in AI.
Vibe coding is becoming one of the most popular use cases for generative services and AI models, driving the valuations of some US companies sky-high. But the Chinese origins of some of the underlying foundation models are raising concerns among observers.
Cognition and SWE 1.5 Suspected of Building on Zhipu AI's GLM Model
Cognition and Cursor, two US companies at the forefront of the vibe coding movement, have introduced their own tools to turn software programming into an agentic process. Both claim to have reached unprecedented performance and efficiency, but neither has disclosed its clear reliance on advanced AI models developed in China.
Cognition recently launched SWE 1.5, its latest AI model, designed and optimized for software engineering. The company says the model was trained on hundreds of billions of parameters and comes close to matching state-of-the-art performance. SWE 1.5 is expected to resolve the tradeoff between speed and code quality that often afflicts AI text generation services.
According to Cognition, its agentic technology uses "a leading base model" released under an open-source license, though no specifics were provided. However, when users asked the tool about its base model, it appeared to reference the GLM family developed by Zhipu AI, a company based in Beijing.
Zhipu AI confirmed the speculation, suggesting that SWE 1.5 may be built on its GLM 4.6 model. Cognition has not commented on the matter. The company is based in the San Francisco area, is valued at 10.2 billion dollars, and aims to deliver the fastest agentic coding experience available on the market.
Anysphere's Vibe Coding and Evidence of Leaked Chinese Text
Anysphere is another AI coding service developer with an extremely high valuation. The company recently introduced Composer, an agent designed to support vibe coding projects in Cursor. The company describes Cursor as the best way to write code with AI, and says Composer can generate code output four times faster than similar models. Anysphere has not disclosed which model Composer uses, but users noticed that it sometimes produces text in Chinese. Like Cognition, Anysphere is based in San Francisco and has tripled its valuation to nearly 10 billion dollars in just six months.
Neither Cognition, Anysphere, nor Cursor has commented on the Chinese origins of the models powering their products. Zhipu AI said that US companies adopting its GLM models is not necessarily a negative, since it can benefit open-source contributions across the AI ecosystem.
According to Florian Brand, a PhD researcher at Trier University in Germany, the origin of the foundation model US companies use is not the most important issue. Building effective AI agents requires an additional stage of fine tuning, which he argues is the real secret behind the success of these projects. Still, this lack of transparency raises a big question about independence and control over core AI technology in the United States.
