ByteDance Reportedly Trains 10 Trillion-Parameter AI Model

ByteDance is reportedly pre-training a huge new artificial intelligence system with as many as 10 trillion parameters.

The project could put the ByteDance AI model among the largest systems developed by a Chinese technology company so far. The Financial Times reported that the company has already started pre-training, citing people familiar with the matter.

ByteDance has not publicly confirmed the report. Reuters said it could not independently verify the details, and the company did not immediately respond to a request for comment.

A model built at enormous scale

Parameters are the numerical settings an AI model learns during training. They help the system identify patterns, generate responses and complete tasks.

A 10 trillion-parameter model would be more than three times larger than Kimi K3, a 2.8 trillion-parameter system from Chinese startup Moonshot AI.

China’s domestic AI sector has already produced several trillion-parameter models. Before Kimi K3 arrived, Meituan’s LongCat-2.0 and DeepSeek’s V4-Pro reportedly led the market with 1.6 trillion parameters each.

However, parameter counts only offer a rough indication of a model’s scale. They do not automatically show how capable, efficient or commercially useful a system will be.

ByteDance could approach leading US systems

The reported size of the ByteDance AI model has drawn comparisons with advanced systems developed in the United States.

Companies such as Anthropic and OpenAI do not disclose the parameter counts behind their most advanced models. That makes direct comparisons difficult.

Still, the Financial Times cited industry estimates suggesting that Anthropic’s Mythos 5 may have around 8 trillion parameters. Estimates put Fable 5 at roughly 5 trillion parameters.

If those estimates are accurate, ByteDance’s model could be close to Mythos 5 in raw scale. The comparison does not confirm that the models will offer similar performance.

Pre-training could take up to six months

The system is reportedly in pre-training, the first major stage of development. During this phase, developers expose a model to large datasets so it can learn language, reasoning patterns and other general capabilities.

According to the report, this process could take between three and six months. ByteDance would then need to fine-tune the system before making it available for products or wider use.

The project highlights the growing pressure on Chinese and US technology companies to develop stronger AI models while controlling the high costs of training and operating them.