China’s ByteDance is Training a 10 Trillion-Parameter AI Model: Report

ByteDance is training an artificial intelligence model with up to 10 trillion parameters, according to a Financial Times report citing people familiar with the matter. 

If completed, it would rank among the world’s largest foundation models.

The TikTok parent is in the early stages of pre-training the model. That phase is expected to take three to six months before fine-tuning begins. The final parameter count has not yet been decided, the report said.

At the reported scale, the model would be about three times larger than Moonshot AI’s Kimi K3, currently China’s largest released model. 

Industry estimates cited by the Financial Times suggest Anthropic’s flagship Mythos 5 contains around eight trillion parameters, though the company has not disclosed official figures.

The report said several Chinese AI labs are building models comparable in size to Anthropic’s Fable 5. ByteDance is reportedly pursuing the largest model among them.

The effort is being led by ByteDance’s Seed AI team, which has around 2,000 researchers, infrastructure engineers, data specialists and translators across China and overseas. The group is headed by former Google DeepMind scientist Wu Yonghui.

Unlike Alibaba, Moonshot and DeepSeek, ByteDance has largely kept its frontier models closed. Its latest 10 trillion-parameter effort suggests the company is prioritising frontier capabilities over open-weight releases.

The report also adds that ByteDance has avoided using model distillation to accelerate development.

Founder Zhang Yiming has instead pushed the company to build models independently, arguing that this is the only path to “world-leading model capabilities”.

In its AI portfolio, ByteDance already offers the Doubao chatbot, China’s most widely used AI assistant with 324 million monthly active users, and the Seedance video generation model. 

It has also expanded its Volcano Engine cloud platform for enterprise AI services while developing its own Seed family of foundation models and investing in custom AI chips.

The latest push comes as Chinese AI companies continue to lead the open-weight model ecosystem. 

Over the past month, Moonshot AI released its 2.8 trillion-parameter Kimi K3 model, while Alibaba unveiled Qwen3.8-Max, a 2.4 trillion-parameter flagship that it plans to release under an open-weight licence. 

Both companies claim performance approaching leading proprietary models from Anthropic and OpenAI, while making their weights available to developers.