Report: ByteDance's Seedance 2.0 is a 200B Parameter Video Model

bdsqlsz · x · 2026-08-08

According to a LatePost interview with former Tencent AI Lab lead Liu Wei, ByteDance's highly acclaimed video generation model, Seedance 2.0, boasts a massive 200 billion parameters.

Liu noted that Seedance 2.0 validates the existence of a scaling law for video models. Its core advantage lies in achieving "unification" through a massive architecture—integrating previously fragmented tasks like text-to-video, multi-image-to-video, and video editing into a single model for multi-task training. He analyzed that Seedance's approach relies on scaling up the model first and feeding it vast amounts of data, essentially betting on brute force.

Liu, who left Tencent in late 2024 to found Video Rebirth (recently raising $80 million), stated he will pursue a differentiated strategy. Facing giant competitors, he aims to achieve equivalent SOTA results using less than one-tenth of Seedance's resources through innovative model architectures and training algorithms.

Related event: ByteDance's Seedance V2 Reported to Have 200B Parameters(2 posts)→

Original post →

More from Companies & People

Companies & People channel →