ByteDance trains an AI model on the scale of Anthropic's Mythos, signaling ambitions to overtake the U.S.

[Financial News]  

ByteDance is currently training a large AI model that appears to be comparable in scale to Mythos, Anthropic's flagship AI model in the U.S. This suggests that China is no longer limited to low-cost AI models that imitate U.S. systems, but is also expanding into frontier AI. The gap with the U.S. is narrowing.

The Financial Times reported on the 7th, citing three sources, that ByteDance has entered the early stage of training a giant AI model capable of handling up to 10 trillion parameters. That would make it three times larger than Kimi K3, the AI model from China's Moonshot that recently jolted the market. Kimi had been the largest AI model ever announced in China, but ByteDance has now begun training one that is three times larger.
According to one source, ByteDance's giant AI model is in the pre-training stage, which typically takes three to six months. After pre-training, it will be released following fine-tuning. The exact model size will be decided later.
Anthropic has not disclosed the size of its Mythos AI models, but industry estimates put the frontier Mythos 5 at about 8 trillion parameters and Fable 4 at about 5 trillion.
However, a larger number of parameters does not automatically guarantee better performance. While parameters are a basic factor that determines a model's information storage capacity, an AI system's actual capabilities are also heavily influenced by other factors, such as the quality of the training data and the training method.

ByteDance's move to train the world's largest AI model shows that Chinese AI companies are not merely following the U.S., but are also trying to outpace it in frontier AI.
Over the past few weeks alone, Chinese AI models from Moonshot and Alibaba Group have demonstrated strong performance on several benchmark tests, falling short only of Anthropic's Fable 5 in certain areas. Mythos 5, which is even more capable than Fable 5, was temporarily barred from deployment in June for security reasons and is now available only to approved organizations, making direct comparison difficult.
Industry insiders said that many Chinese companies are already training AI models at the scale of Fable 5, and that ByteDance is pursuing its ambition to build the largest possible AI model.

ByteDance is taking a different path from other Chinese AI companies. While many Chinese tech giants have adopted open-source approaches that release their code, ByteDance keeps its code closed. At the same time, it is known for strong performance.
ByteDance's latest Seedance model is regarded as the world's best-performing model in video generation. Its flagship consumer AI service, Doubao, has 324 million monthly active users, making it the most popular AI platform in China.
Over the past three years, ByteDance has invested more aggressively in AI than any other major Chinese tech company, building a large-scale data center network and hiring researchers in large numbers. It is also focusing on Volcano Engine, its cloud unit that provides AI services to businesses, and has ambitions to develop custom AI chips.
The model development team Seed, led by Yonghui Wu, a scientist from Google DeepMind, has about 2,000 employees in China and overseas. The team includes core researchers, infrastructure engineers, data labeling staff and translators.

According to one source, ByteDance's AI model development is not based on so-called distillation, a method that extracts the core elements from U.S. AI models. Instead, it is pursuing an independent in-house model development approach. ByteDance has stuck with this method for more than a year, and as a result, its model development is said to have been slower than that of Chinese rivals.
[email protected] Song Kyung-jae Reporter