According to media reports, ByteDance is discussing the training of a model with more than 5 trillion parameters - a number surpassing Alibaba's Qwen 3.8-Max with 2.4 trillion and Moonshot's K3 with 2.8 trillion, making it the largest known parameter model in China so far. The plan is still in its early stages, and the model may not be released in the end. The new model is led by Xiang Liang, head of Seed Foundation, and co-led by Shen Ke, who is responsible for pre-training data of large language models. Seed is reorganizing itself, defining responsibilities, and allocating resources for this project. A person close to Seed gave a straightforward reason: rather than continuing to catch up with other large models in existing sizes, they would push the parameter scale to several times that of their peers to secure a leading position.
This is backed by a series of clear statements from ByteDance's senior leadership. Founder Zhang Yiming made a rare statement at a recent internal meeting: "We should persist in long-termism and delayed gratification when building models, rather than using others' outputs to chase short-term rankings." He clearly stated that ByteDance should be willing to sacrifice some short-term gains for long-term goals. Even harsher is the current strict ban on distilling open-source models within ByteDance, with even API detection methods used to strengthen restrictions - Zhang Yiming believes that distillation interferes with true long-term technological breakthroughs.
At the company-wide meeting on August 6, 2026, CEO Liang Rubo reiterated this logic again. He said that ByteDance will continue to focus on self-research and basic skills in large language models, accept short-term setbacks, and persist in optimizing the long term. By optimizing the long term, it means focusing on user value and incremental social value, not being swayed by short-term fluctuations; in terms of AI business, it means believing in AGI and sticking to self-research. He specifically clarified, "Yesterday's external report said that we are insisting on self-research due to external pressure, which is pure speculation. The real reason is that we want the team to have delayed gratification, build solid technical foundations, which is crucial for achieving AGI in the long run."
From Zhang Yiming to Liang Rubo, this series of high-level statements is not just a technical plan but a complete strategic declaration: to unify R&D thinking internally and send signals externally - ByteDance has no intention of following others' technology routes in a vicious cycle and plans to chart its own growth curve.
