ByteDance is training a 10 trillion parameter model and says it won't copy the west to do it

The company that owns TikTok is attempting to build the largest AI model China has ever produced, and it has reportedly been told to do it without copying anyone.

The Scale

According to reporting by the Financial Times on August 7, citing three people familiar with the project, ByteDance is pre-training a model with as many as 10 trillion parameters.

For comparison, Moonshot AI's Kimi K3 — released as open weights on July 27 and described at the time as the largest open-weight model in history — has 2.8 trillion parameters. Alibaba's Qwen3.8-Max has 2.4 trillion. ByteDance's model would be more than three times either of them, and would place it among the largest systems being built anywhere in the world.

The model uses a mixture-of-experts architecture, meaning only a fraction of its parameters activate for any given input. The effort is reported to require approximately 30,000 GPUs and three to six months of continuous pre-training, run by ByteDance's Seed AI team of around 2,000 staff.

That timeline is the detail worth marking: three to six months from now places results, or problems, before the end of 2026.

The Instruction Not To Distill

The most consequential element of the report is not the parameter count. It is that founder Zhang Yiming has reportedly directed the Seed AI team to pursue genuine technical breakthroughs rather than relying on distillation from Western AI systems.

Distillation — training a model to imitate the outputs of a more capable one — is precisely what the White House accused Moonshot AI of doing last month. On July 22, science and technology policy director Michael Kratsios publicly alleged Moonshot had conducted large-scale covert distillation of Anthropic's Fable model to build Kimi K3. Treasury Secretary Scott Bessent warned that ‘sanctions and Entity List designations will be on the table,’ writing that ‘open source is not open season on American IP.’

Against that backdrop, an explicit internal directive to avoid distillation reads as a strategic position rather than a technical preference. A Chinese laboratory building a frontier model entirely from its own training runs is considerably harder to accuse of theft — and considerably harder to sanction.

Whether the instruction is followed, and whether outside parties accept that it was, are separate questions that a finished model's weights would help answer.

Why ByteDance Specifically

ByteDance is not the obvious candidate for China's largest model. DeepSeek, Alibaba, Moonshot and Zhipu have been the names on leaderboards — in mid-July, six of the top ten models on independent rankings were Chinese, including all of the top five.

But ByteDance has two advantages that matter at this scale. It has data at a volume few companies anywhere can match, from a recommendation system serving over a billion users. And it has the capital to commit 30,000 GPUs to a single training run for half a year without needing that run to produce revenue.

The 2,000-person Seed AI team is also unusually large for a single research organisation, comparable to the headcount of entire frontier laboratories in the United States.

What It Signals

The broader significance is about where scale competition sits now. Western laboratories have moved noticeably toward efficiency — smaller active-parameter counts, aggressive price cuts, distillation of their own large models into cheaper ones. OpenAI cut prices up to 80 per cent on its lower tiers in July. DeepSeek's V4 Flash reaches near-frontier coding performance while activating only 13 billion of its 284 billion parameters.

ByteDance is going the other way, betting that raw scale still has room. Ten trillion parameters is a wager that the returns to size have not yet flattened — which, if it pays off, resets the competition for everyone, and if it does not, will have been an expensive six months of GPU time.