Zhipu Founder Outlines Next-Gen GLM-5.4/5.5 Plans: Trillion-Parameter Scale, RSI and Agent Focus

At a speech in Singapore on October 8, Zhipu founder Tang Jie outlined the R&D direction for the company's next-generation large models, GLM-5.4/5.5. He said the new models will significantly expand pretraining data and scale, targeting several T, with parameters rising from the current 740+ billion level to at least 1 trillion, aiming to raise the upper bound of model capability.

In post-training, Zhipu will continue to strengthen reinforcement learning methods such as SFT, RLHF, and RLVR. The models are also expected to improve long-horizon reasoning and agent abilities, handling more complex environments, harder tasks, and longer cycles, giving models more room for thinking and operation as they move toward long-horizon tasks and autonomous agents.

Tang said GLM-5.4/5.5 will follow two new paths: RSI (recursive self-improvement) for autonomous iterative upgrades, and longer, more autonomous agents. At an earnings call in late August 2026, he had described the next-generation large model as fully self-training, meaning the model could train itself from pretraining through post-training and evolve autonomously. According to related interpretations, GLM-5.4/5.5 is unlikely to achieve this fully, but may incorporate some RSI capabilities, with fully autonomous RSI evolution still some way off.