The Chinese technological giant ByteDance is testing a new artificial intelligence model (IA) that could reach the scale of the most advanced systems in the U.S. Anthropic, today reported the Financial Times (FT).
© Lusa
08/08/2026
by
Lusa
A owner of the TikTok platform is in the early stage of pre-development of a model that can reach up to 10 billion parameters, quoted the British daily, which heard three sources aware of the process, but which were not identified.
The new ByteDance system will be three times larger than Moonshot's Kimi K3, currently the largest AI model developed in China.
The number of parameters defines the fundamental capacity and memory limits of a model to store information, although the actual efficiency also depends on factors such as data quality and training methods.
Anthropic does not disclose the size of its models, but sector estimates indicate that the most advanced system, Mythos 5, has about eight billion parameters, while Fable 5 is about five billion.
According to the sources cited by France Télécom, the pre-development process of ByteDance should be extended by three to six months, prior to the fine tuning phase and the eventual launch, and the exact size of the model will only be fixed at a later stage.
ByteDance's initiative reflects Chinese laboratories' ambition to shorten the distance for US leaders and compete at the top of the global artificial intelligence sector.
In recent months, several Chinese models developed by companies such as Moonshot and Alibaba have demonstrated competitive performances in benchmark tests, approaching the levels of Fable 5.
Unlike Chinese competitors who chose to publish open source models, ByteDance has maintained a more reserved approach. Its SeeDance video generation model is among the most advanced worldwide, while Doubao chatbot leads the Chinese market with 324 million monthly active users.
Led by founder Zhang Yiming, ByteDance has strengthened investments in data centers, semiconductors, and the recruitment of researchers for the AI development team, named Seed and composed of about 2,000 professionals worldwide.
The AI security controls that the U.S. Government intends to apply will target the closed models of OpenAI, Google and Anthropic, saving open Meta or DeepSeek models, reported this week ‘media’ Americans.
Closed AI models remain housed on their creator's servers, which maintain control over their use without revealing the programming. On the contrary, the user of an open AI model can download and modify it.