ByteDance Discusses 5T Parameter Model, Surpassing Qwen3.8-Max and K3 in Size

ByteDance explores training a 5 trillion parameter model, exceeding Alibaba’s Qwen3.8-Max and Dark Matter’s K3. Led by Xiang Liang and Shen Ke, the project remains in early stages with no release guarantee.

Woofun AI reports that ByteDance is evaluating the development of an AI model featuring more than 5 trillion parameters. This scale would exceed Alibaba’s Qwen3.8-Max, which holds 2.4 trillion parameters, and Dark Matter’s K3, with 2.8 trillion parameters, positioning it as the largest known model in China by parameter count. The initiative is spearheaded by Xiang Liang, head of Seed Foundation, alongside Shen Ke, head of Large Language Model Pre-training Data. To support this effort, Seed is restructuring its operations to clarify responsibilities and allocate resources. The project remains in preliminary discussions, with no assurance of a final release.