ByteDance Trains AI Model With Up to 10 Trillion Parameters to Rival Anthropic
ByteDance is training a model with up to 10 trillion parameters, aiming to surpass Anthropic as Chinese AI labs narrow the gap with leading US rivals.
Summary
On August 7, 2026, the Financial Times reported, citing three people familiar with the project, that ByteDance had begun pretraining a model with as many as 10 trillion parameters. That is three times Moonshot’s Kimi K3, the largest released Chinese model. Pretraining typically takes 3 to 6 months, followed by fine tuning and a release if successful, while the final size remains undecided. Industry estimates put Anthropic’s Mythos 5 at about 8 trillion parameters and Fable 5 at 5 trillion. Parameter count sets capacity and memory limits, but performance also depends on data quality and training methods.
Recent Moonshot and Alibaba models trail only Fable 5 in some benchmarks, and several Chinese labs are training similarly sized systems, although ByteDance has the largest stated target. Mythos 5 remains limited to approved organizations following a temporary June security ban. The TikTok parent keeps most models closed, but SeeDance ranks among the leading video generators and Doubao leads China with 324 million monthly active users. Over three years, ByteDance has outspent other Chinese technology giants on AI, data centers and researchers, expanded enterprise AI cloud unit Volcano Engine, and pursued custom chips.
Seed, led by former Google DeepMind scientist Wu Yonghui, has about 2,000 researchers, infrastructure engineers, data labelers and translators in China and overseas. For more than a year it has rejected distillation, where smaller models copy larger teachers, an independent strategy some believe slowed progress. Founder Zhang Yiming told Seed two weeks earlier to pursue world-leading long-term capabilities despite near-term delays. Latepost and The Information first reported his remarks. ByteDance did not respond to a comment request.
Positives
- Up to 10 trillion parameters would make ByteDance’s model three times larger than Moonshot’s Kimi K3.
- 324 million monthly active users make Doubao China’s most popular consumer AI model.
- SeeDance already ranks among the world’s most advanced video generation models.
- About 2,000 Seed personnel support research, infrastructure, data labeling and translation across China and overseas.
- Three years of aggressive investment have expanded ByteDance’s data centers, research workforce, Volcano Engine cloud business and custom chip ambitions.
Risks & concerns
- The final parameter count remains undecided, and release depends on successful pretraining and fine tuning after a typical 3 to 6 month initial stage.
- Parameter count measures capacity limits, not capability, which also depends on data quality and training methods.
- More than a year without model distillation may have slowed ByteDance’s development relative to Chinese rivals.
- ByteDance keeps most of its models closed, unlike many competing Chinese AI laboratories.
- Anthropic restricted Mythos 5 to approved organizations after a temporary June ban prompted by security concerns.