ByteDance is training an AI model that could approach the size of Anthropic's most advanced system, Mythos, the Financial Times reported on August 7, citing people with knowledge of the matter.
Industry estimates cited in the report suggest the new model could reach up to 10 trillion parameters — roughly three times the size of ByteDance's previous largest model. For context, Anthropic's most advanced Mythos 5 is estimated at about 8 trillion parameters, while Alibaba's Fable 5 is estimated at around 5 trillion. Anthropic does not disclose the sizes of its models, and Reuters said it could not immediately verify the FT report.
The scale-up is the latest sign of an intensifying global race to build ever-larger frontier models, with Chinese companies increasingly targeting the same scale as their US rivals. ByteDance — the parent of TikTok and Douyin — has pushed aggressively into AI with its Seed research team, open-source Qwen rival models and the Seedance video-generation line, which the company says maintains state-of-the-art status in global video generation.
The pursuit of a ~10-trillion-parameter model would put ByteDance in rarefied company. Training such models requires enormous computing power — and, given ongoing export restrictions on advanced AI chips to China, raises open questions about how the company would source the compute. It also comes as Chinese labs pursue a strategy of efficiency alongside scale, with models like DeepSeek demonstrating strong performance at far lower cost.
If the report is accurate, ByteDance's move underscores that despite export controls, compute constraints and efficiency breakthroughs, China's biggest AI players are still betting that raw scale remains the path to frontier capability.
