[Zhipu responds to the question of "knowledge only after training": The next generation base model is already advancing GLM-6.0, aiming at "self-evolution"] September 1st, at the 2026 semi-annual performance communication meeting held on the evening of August 31, an analyst raised the market's doubts about Zhipu's "knowledge only after training" and asked why the company has been relatively restrained in parameter scale recently. Tang Jie, founder of Zhipu, said that the next generation model will still expand the base size, and at the same time control the activation parameters to avoid the decline in inference speed and increase in costs. In terms of computing power, Zhipu revealed that it has achieved large-scale inference with 100,000-level domestic chips, and the cost of inference per Token has dropped by 80% from the beginning of the year. GLM‑5.3‑Flash is Zhipu’s first model that fully relies on domestic chip clusters to provide services under ultra-large-scale real traffic. The company said that compared with the initial baseline of the same hardware, the end-to-end service performance of this model has been improved by 3 times. Tang Jie summarized one of the future directions of GLM-6.0 as self-evolution, "This is a big challenge. The biggest problem is that the model can judge by itself when to stop and when to correct itself. This is the biggest problem, rather than simply training the model to a larger size. In future models, this is a key research focus." (every sutra)