ByteDance's Seedance V2 Hits 200B Parameters, Next-Gen Could Reach 1T

Scobleizer · x · 2026-08-07

According to leaks, ByteDance's video generation model Seedance V2 is a 200B parameter MoE architecture model, and the upcoming V2.5 will be even larger. Seedance V3 is expected to become the first 1T parameter video model next year. Additionally, ByteDance's language model Dola Seed may reach around 5T parameters in the near future.

Citing a LatePost report, ByteDance continues its resource-heavy approach in AI. While many teams believe scaling traditional DiT architectures yields diminishing returns, Seedance invested heavily in pushing pre-training to the limit. This strategy led V2 to the top of the field. Scaling model size has become an industry consensus, with companies like Moonshot and xAI also racing to build ultra-large models.

Original post →

More from Companies & People

Companies & People channel →