ByteDance is reportedly training a 10-trillion-parameter AI model

ByteDance is pre-training an artificial-intelligence model with as many as 10 trillion parameters, according to people familiar with the effort, a scale that would approach Anthropic’s largest Mythos systems.

If accurate, the model would be more than three times the size of Moonshot AI’s Kimi K3, which carries about 2.8 trillion parameters, and would rank among the largest known training runs by a Chinese company.

Parameter count is a rough proxy for capability and cost, not a guarantee of either. Larger models demand vastly more compute and data, and recent frontier work has leaned on efficiency and mixture-of-experts designs as much as raw size. A run at this scale would still signal ByteDance’s willingness to spend at the frontier even as US export controls limit Chinese access to top-tier chips.

Industry estimates put Anthropic’s Mythos 5 at roughly 8 trillion parameters and its Fable 5 at about 5 trillion, though Anthropic does not disclose parameter counts for any model. Those figures are outside estimates, not company-confirmed numbers.

The model is said to be in pre-training, a phase that typically runs three to six months before fine-tuning and any release decision, and the final parameter count has reportedly not been locked in. ByteDance did not respond to a request for comment, and the account could not be independently verified.

Even a completed run would not settle how the model performs; scale sets a ceiling, not a score.