ByteDance is reportedly in discussions to train an artificial intelligence model with over 5 trillion parameters, according to a recent report from LatePost.

This proposed model would surpass Alibaba's Qwen 3.8-Max, which has 2.4 trillion parameters, and Moonshot AI's K3, which features 2.8 trillion parameters. If developed, it would become the largest known model in China by parameter count. The plan is still in its early stages, and the report notes that the model may not ultimately be released.

The new model would be led by Xiang Liang, head of the Seed Foundation, in collaboration with Shen Ke, who leads pre-training data for the large language model. In response, the Seed team is reorganizing its structure, clarifying responsibilities, and reallocating resources. Both Xiang Liang and Shen Ke are key figures within ByteDance’s AI and large model development framework, with backgrounds in the company's "search, advertising, and recommendation" system.

Xiang Liang earned his undergraduate degree from the University of Science and Technology of China and a PhD from the Institute of Automation, Chinese Academy of Sciences. He joined ByteDance in 2016, previously leading the AML machine learning platform team, before moving to lead the Doubao large model Foundation team. Shen Ke graduated from Tsinghua University in 2018, joined ByteDance, and now primarily oversees LLM pre-training data.

Regarding the rationale for training a model with over 5 trillion parameters, the report, citing sources close to the Seed team, suggests that the team believes it is more effective to scale the model's size to several times that of competitors to secure a leading position, rather than continuing to catch up on existing model sizes.