ByteDance Begins Training a 10-Trillion-Parameter AI Model, Approaching Anthropic’s Scale

According to Beating Monitor, the Financial Times, citing three informed sources, reported that ByteDance has begun pre-training a large model with up to 10 trillion parameters. The project is still in its early stages, and the final scale has not been determined. If trained to the upper limit, it would exceed Kimi K3 by more than three times, becoming the largest model by parameter count known to be developed by a Chinese team. ByteDance is now directly pushing its model size into the vicinity of Anthropic’s flagship models. Although Anthropic has never disclosed the parameter count for Mythos 5, the Financial Times cited industry estimates suggesting Mythos 5 is around 8 trillion parameters and Fable 5 around 5 trillion. If ByteDance’s new model reaches 10 trillion parameters, it would at least enter the same parameter scale as America’s most advanced closed-source models. The model is currently in the pre-training phase (training the model’s foundational capabilities on massive datasets). The Financial Times noted that this phase typically takes three to six months, followed by additional post-training before a release is possible. More parameters do not necessarily equate to stronger performance; final capabilities depend on architecture, training data, and training methods. Zhang Yiming recently made it clear internally at Seed that he opposes distilling competitors’ models and has instructed his team to accept short-term lagging behind while pursuing world-leading status through independent R&D. This new model, with up to 10 trillion parameters, represents ByteDance’s most aggressive bet on scale to date.