Zhipu’s New Model Reportedly Features Up to 3.5 Trillion Parameters, Potentially Reaching Claude Fable-Level Performance

According to a Beating AI bulletin, an expert interview transcript circulating among investment circles reveals that Zhipu AI’s upcoming next-generation foundation model will have a parameter count ranging from 2.5 trillion to 3.5 trillion—not the previously rumored 1.6 trillion. Meanwhile, GLM-5.3, released by Zhipu in August this year, is still built upon the prior-generation foundation model and has approximately 744 billion parameters. If accurate, the new foundation model’s parameter count would be 3.4x to 4.7x that of GLM-5.3. The transcript notes that the model’s training encountered setbacks, prompting Zhipu to adopt a more aggressive architectural design—leading to prolonged stagnation of the loss function during pre-training, i.e., sustained compute consumption without meaningful improvement in key training metrics. However, the interviewee assesses that these technical issues have recently been resolved, with training expected to wrap up imminently; the final model is anticipated to reach flagship-tier performance comparable to Claude Fable. The transcript also states that the new model’s…