Zhipu's new model reportedly has up to 3.5 trillion parameters, potentially reaching Claude Fable level.

Beating AI News Flash: A summary of an expert interview circulating in the investment community claims that Zhipu's next-generation foundation model is not the previously rumored 1.6 trillion parameters, but has reached 2.5 trillion to 3.5 trillion parameters.

GLM-5.3, released by Zhipu in August this year, still uses the previous-generation foundation, with total parameters of approximately 744 billion. If the information is accurate, the scale of the new foundation will expand to about 3.4 to 4.7 times that of the current version.

However, the training process was not smooth. The summary states that Zhipu this time adopted a more aggressive model architecture, and during pre-training there was a period when Loss did not decrease, meaning computing power continued to be invested but training metrics could not continue to improve.

The interviewee believes the problem should have been resolved recently, expects training to be completed soon, and predicts that the new model can reach the flagship level of Claude Fable.

The new model also aims to solve the problem of large models being "releasable but unable to meet supply" after launch. The summary states that Zhipu has treated inference speed and cost as hard constraints since the architecture design stage, with the goal of being able to provide large-scale service on the first day of release, and that multiple cloud vendors have already discussed revenue sharing.