ByteDance Reportedly Training 10T Parameter Model; Zhang Yiming Rejects Distillation Shortcuts

智东西 · wechat · 2026-08-07

ByteDance is reportedly training a massive AI model with an estimated 5 to 10 trillion parameters, led by Seed Foundation head Xiang Liang. The model is currently in the pre-training phase and aims to rival or surpass leading frontier models in scale.

During an internal meeting, founder Zhang Yiming explicitly spoke out against model distillation. He argued that while distillation improves short-term performance, it essentially copies existing capabilities and hinders genuine technological breakthroughs. He emphasized that ByteDance would not take this shortcut, even if it meant temporarily lagging behind domestic competitors. Meanwhile, leadership acknowledged that the Doubao model currently lags in AI Coding, an area they plan to heavily prioritize alongside continued massive investments.

As overseas frontier models face safety setbacks, Chinese AI companies are seizing a critical window of opportunity, with ByteDance aggressively pushing both massive model training and B-side product integration.

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

Original post →

More from Companies & People

Companies & People channel →