Zhipu AI's new model has been leaked to feature up to 3.5 trillion parameters, and is expected to reach the performance level of Anthropic's Claude Fable.

Beating AI Express (Dongcha) reports that an expert interview transcript circulating in investment communities claims Zhipu AI’s next-generation base model does not carry the previously rumored 1.6 trillion parameters, but ranges from 2.5 trillion to 3.5 trillion parameters. Zhipu’s GLM-5.3, released in August this year, still uses the previous-generation base model, with around 744 billion parameters. If the news is accurate, the new base model’s scale will expand to roughly 3.4 to 4.7 times that of the current version. However, the training process has not been smooth. The transcript states Zhipu adopted a more aggressive model architecture this time, leading to a period during pre-training where the loss did not decrease – meaning continuous computing power investment failed to improve training metrics. The interviewee believes the issue has likely been resolved recently, with training expected to wrap up soon, and predicts the new model will reach the flagship level of Anthropic’s Claude Fable. The new model also aims to solve the common problem of large language models being "launchable but not supply-ready" after release. The transcript notes Zhipu has set inference speed and cost as hard constraints from the architectural design phase, targeting large-scale service availability on launch day. Multiple cloud vendors are already in discussions over revenue sharing.