ByteDance Training AI Model with 10 Trillion Parameters: Report

TikTok's parent company, Chinese tech firm ByteDance, is training a large-scale artificial intelligence (AI) model that could reach up to 10 trillion parameters, the UK daily Financial Times (FT) reported on August 7 (local time).

If these projections hold, the model under development by ByteDance would be comparable in parameter count to US AI firm Anthropic's high-performance advanced model, Mythos.

The largest Chinese AI model unveiled to date was Moonshot's Kimi K3, which has 3 trillion parameters.

The ByteDance model currently under development is undergoing pre-training, which typically takes three to six months, and is scheduled to be released following fine-tuning if everything goes smoothly.

The exact size of the model has not yet been finalized.

While Anthropic does not disclose the exact number of parameters used in its models, industry estimates suggest that its latest advanced model, Mythos 5, has about 8 trillion parameters, while Fable 5, which offers slightly lower performance, has about 5 trillion.

The performance of an AI model depends not only on the number of parameters, but also on other factors such as the quality of the training data and the training methodology.

In recent weeks, AI models released by China's Moonshot and Alibaba have received favorable reviews in comparative tests, with some areas trailing only Anthropic's Fable 5.

Anthropic's latest advanced model, Mythos 5, is currently available only to select customers.

While several research and development teams in China are reportedly training models on the scale of Fable 5, industry insiders told the FT that ByteDance is the most ambitious in attempting to build a model on par with Mythos 5.

Notably, unlike many Chinese companies that lean toward open-source AI models, most of ByteDance's models are closed-source.

ByteDance's "Seedance" is evaluated as one of the world's most advanced video generation models, and Doubao, which is available to general consumers, is the most popular AI model in China with 324 million monthly active users.

Over the past three years, ByteDance has invested more aggressively in AI than any other Chinese tech giant, expanding its data centers and hiring researchers.

ByteDance has also increased investments in its cloud business unit, Volcano Engine, which sells AI solutions to enterprises, and is pursuing the development of its own AI chips.

In early 2023, ByteDance established Seed, a basic AI model research team, and Wu Yonghui, who worked at Google and its subsidiary DeepMind before joining ByteDance in February 2025, currently serves as the research head of the Seed team.

According to the FT, the Seed team currently employs about 2,000 people across China and overseas, including core researchers, infrastructure engineers, data labeling teams, and translators.

A person familiar with the matter told the FT that the ByteDance Seed team has chosen an approach to develop models more independently, rather than relying on the "distillation" of existing models from other labs.

This approach has been applied for over a year, and some believe this is why their development pace may appear slower than that of competitors.

"Distillation" refers to the process of training a smaller model to mimic and compress the knowledge and output results of a larger, more complex AI model.

However, ByteDance founder Zhang Yiming is known to be convinced that surpassing competitors can only be achieved through independent development rather than distillation.

Zhang reportedly told the Seed team during an internal meeting about two weeks ago not to be overly concerned about lagging behind in the short term, and to aim for "world-class model capabilities" in the long run.

These remarks were reported by Chinese economic and tech media outlet LatePost and US information technology media outlet The Information.

ByteDance did not respond to a request for comment, the FT reported.
(Photo: Getty Images)
※ Please note: This article was translated by AI and may contain errors.