According to a report by the Financial Times on Friday, citing people familiar with the matter, the TikTok parent company is training a new AI model with as many as 10 trillion parameters, a scale that could place it among the world's most powerful AI systems.
If confirmed, the model would significantly surpass the size of existing Chinese large language models. It would be more than three times larger than Moonshot AI's Kimi K3, which has 2.8 trillion parameters, and well ahead of other leading domestic models.
In artificial intelligence, parameters are the numerical values a model learns during training. They enable an AI system to recognize patterns, process information, generate responses and perform a wide range of tasks. While parameter count is often used as a measure of a model's scale, researchers note that larger models are not automatically more capable, as overall performance also depends on factors such as training data, architecture and optimization.
Before the launch of Kimi K3, China's largest AI models included Meituan's LongCat-2.0 and DeepSeek's V4-Pro, each with around 1.6 trillion parameters. Several other Chinese developers have also crossed the one-trillion-parameter milestone as competition in the country's AI sector accelerates.
Comparing Chinese models directly with those developed by U.S. companies remains difficult because leading American AI firms, including Anthropic and OpenAI, do not publicly disclose the parameter counts of their flagship systems, such as Mythos, Fable, and GPT-5.5.
However, the Financial Times reported that industry estimates place Anthropic's flagship Mythos 5 model at roughly 8 trillion parameters, while Fable 5 is believed to have around 5 trillion parameters. If ByteDance's reported 10-trillion-parameter model reaches completion, it would be comparable in scale—and potentially even larger—than Anthropic's most advanced system.
The report said the ByteDance model is currently in the pre-training phase, one of the most computationally intensive stages of AI development. This process typically takes between three and six months, during which the model learns from vast amounts of data before undergoing fine-tuning for specific tasks and eventual public or commercial deployment.
The reported project highlights the increasingly fierce competition between Chinese and U.S. technology companies to develop more advanced artificial intelligence systems. Alongside improving performance, developers are also racing to reduce computing costs and make next-generation AI models more efficient to train and operate.
With governments and technology firms investing billions of dollars into AI infrastructure, ByteDance's reported effort signals China's determination to remain competitive in the global race to build frontier AI models capable of rivaling the industry's biggest players.