ByteDance Reportedly Training Up to 10-Trillion-Parameter Model, Rejecting Distillation

APPSO · wechat · 2026-08-07

ByteDance is training a massive foundation model potentially reaching 10 trillion parameters, currently in early pre-training, according to the Financial Times. If confirmed, it would be China's largest known model, rivaling top global labs like Anthropic.

ByteDance's Seed team comprises about 2,000 employees across the full R&D stack. Founder Zhang Yiming recently told staff that the company will not use "distillation" from leading models as a shortcut, insisting on an independent training path. As the gap with frontier models narrows, ByteDance is betting heavily on raw Scaling to push capability boundaries.

Related event: ByteDance reportedly pretraining 5-10T-param model; Zhang Yiming opposes distillation(19 posts)→

Original post →

More from Companies & People

Companies & People channel →