AIBRIEF
The 10 trillion parameter target, roughly three times Chinese AI lab Moonshot's released Kimi K3 model and near industry estimates for Anthropic's top model, is a capacity ceiling, with the initial training phase, three to six months from completion.
Chinese tech company ByteDance is pre-training an AI model targeting up to 10 trillion parameters, a scale roughly three times the largest released Chinese model and close to industry estimates for Anthropic's most advanced system.
Parameters are a model's learned connections; more of them raise the memory and compute ceiling, but do not by themselves determine capability. Data quality, training method, and post-training matter at least as much. The 10 trillion figure is also a target, not a final size, and the model is in pre-training, the longest stage of building a frontier model, typically three to six months. The actual parameter count is not yet locked.
For comparison, Moonshot's Kimi K3, the largest released Chinese model, sits at roughly 2.8 trillion parameters. Anthropic, the US lab ByteDance is reportedly chasing, does not disclose model sizes; industry estimates place Mythos 5 at about 8 trillion and Fable 5 at about 5 trillion. Mythos 5 is currently restricted to approved organizations after a temporary June ban over security concerns.
ByteDance is one of several Chinese labs reportedly training Fable-5-scale systems. Recent releases from Moonshot and Alibaba already trail only Anthropic's Fable 5 on some benchmarks. ByteDance declined to comment.
The next signal lands when pre-training ends, three to six months from now.