ByteDance is pre-training a 10-trillion-parameter model — three times Kimi K3, FT reports
ByteDance is training a model with as many as 10 trillion parameters, the Financial Times reported Friday — more than three times the size of Moonshot's 2.8-trillion-parameter Kimi K3, and close to industry estimates for Anthropic's Mythos 5. The model is in pre-training, a phase that typically runs three to six months, and its final size will be settled later; ByteDance did not comment.
Why it mattersUntil now China's frontier story has been efficiency — smaller, cheaper, open. A TikTok-scale company betting on raw Mythos-class scale is a different strategy, and ByteDance has the compute, cash and data to attempt it despite US chip restrictions. If the run succeeds, the gap the US labs still hold at the very top gets contested for the first time.
Read in the app — free, in 9 languages
Related stories
ChatGPT is signing its fake cartoons with real cartoonists' names.2026-10-06Qualcomm will license Huawei's way of stacking two chip wafers.
2026-10-06Wikipedia's owner says OpenAI agents may have caused a May outage.
2026-10-06OpenAI fired three safety staff for talking to an outside watchdog.
2026-10-05OpenAI will put ads beside the pictures you ask ChatGPT to draw.
2026-10-05