ByteDance is training an AI model whose scale could approach Anthropic's most advanced Mythos system, the Financial Times reported, as Chinese companies continue to narrow the gap with top U.S. AI labs.
Three people familiar with the matter said the Chinese tech giant is in the early stages of training an ultra-large-scale model that could reach 10 trillion parameters — roughly three times the size of Kimi K3, the largest model released so far in China, developed by Moonshot AI.
One of the sources said the model is currently undergoing pretraining, a phase that typically lasts three to six months. If all goes well, fine-tuning and a release would follow. The exact scale is expected to be finalized at a later stage.
Anthropic has not publicly disclosed its model sizes. Industry estimates, however, indicate that its most advanced Mythos 5 model carries around 8 trillion parameters, while Fable 5 has about 5 trillion.
It is worth noting that parameter count dictates a model's base capacity for storing information, or its memory ceiling — but real-world performance also depends on other factors such as data quality and training methodology.