Zhipu AI's new model is rumored to reach 3.5 trillion parameters, comparable to Claude Haiku.

ChainCatcher report, according to Beating AI’s breaking news, an expert interview summary circulating in investment circles states that Zhipu AI’s next-generation foundation model is not the previously rumored 1.6 trillion parameters, but rather between 2.5 and 3.5 trillion parameters. Zhipu’s GLM-5.3, released in August this year, still uses the previous-generation foundation model with approximately 744 billion total parameters; if the report is accurate, the new foundation model would be 3.4 to 4.7 times larger than the current one. The summary also indicates that training has not gone smoothly—Zhipu adopted a more aggressive model architecture, during pre-training which at one point saw loss values fail to decrease, meaning continued computational investment yielded no further improvement in training metrics. The interviewee believes the issue has likely been resolved recently and expects training to be completed soon, predicting the new model will reach flagship-level performance comparable to Claude Haiku. The summary further notes that from the initial architecture design phase, inference speed and cost were treated as hard constraints, with the goal of enabling large-scale service deployment on the day of launch. Meanwhile, multiple cloud providers have already begun discussions on revenue-sharing arrangements to address the challenge of high demand exceeding supply.