ByteDance is training an artificial intelligence model with as many as 10 trillion parameters, a system that could approach the scale of Anthropic's cutting-edge Mythos model and dramatically raise the stakes in the global AI race. The Financial Times reported the plan on August 7, 2026, citing people with knowledge of the matter, with Reuters and other outlets subsequently confirming the details.
The build is the latest signal that Chinese technology firms are refusing to concede the frontier-model race to their American rivals. For continuous breaking AI news and analysis of where each lab stands, AI Buzz Wire is tracking this story as it develops.
What the 10-Trillion-Parameter Figure Means
A model's parameters are the numerical weights it learns from training data to recognize patterns, generate text, and carry out tasks. They are widely used as a rough proxy for scale, though researchers caution that parameter count alone does not determine how capable a model is. Training methods, data quality, and architecture matter just as much.
Even so, a 10-trillion-parameter target is significant. At that size, ByteDance's model would be more than three times larger than Moonshot AI's Kimi K3, which counts 2.8 trillion parameters and currently represents one of the most ambitious Chinese efforts to date.
Before Kimi K3's release, Meituan's LongCat-2.0 and DeepSeek's V4-Pro led China's domestic field with roughly 1.6 trillion parameters each, while several other Chinese labs had crossed the trillion-parameter threshold. ByteDance's planned system would leave all of them behind by a wide margin.
Closing the Gap With Anthropic's Mythos
Direct comparisons with leading American models are difficult because companies such as Anthropic and OpenAI do not publicly disclose the parameter counts for systems including Fable, Mythos, or GPT-5.5. However, Financial Times cited industry estimates putting Anthropic's most advanced Mythos 5 at roughly eight trillion parameters and Fable 5 at about five trillion.
That would place ByteDance's forthcoming model close to the size of Mythos, the system widely regarded as among the most powerful frontier models available. Reaching that scale would mark a milestone for a Chinese lab and intensify questions about whether the parameter gap between US and Chinese frontier models is finally narrowing.
ByteDance did not immediately respond to requests for comment, the reports noted.
Pre-Training Is Already Underway
The ByteDance model is currently undergoing pre-training, according to the Financial Times. Pre-training is the most compute-intensive phase of developing a large model, and the process typically lasts three to six months before a system can be fine-tuned and released.
That timeline suggests any public release is still months away, though the decision to begin pre-training at this scale reflects a serious, well-resourced commitment. Pre-training a 10-trillion-parameter model requires enormous clusters of specialized chips and vast quantities of electricity, costs that only a handful of companies in the world can absorb.
ByteDance, best known as the parent company of TikTok and Douyin, has previously outlined artificial intelligence as one of its four strategic priorities for 2026. The company has been pushing into generative AI through its Doubao chatbot and Seedance video model, but a frontier-scale language model would represent its most ambitious effort yet to compete head-to-head with the leading labs.
Why the Parameter Arms Race Matters
The reported push comes as Chinese tech firms continue to accelerate their model release cycles, battling US rivals to build more powerful systems without making them prohibitively expensive to run. The strategy has produced a flurry of releases throughout 2026, from Moonshot's Kimi K3 to Alibaba's Qwen3 line and DeepSeek's V4 series.
At the same time, the broader policy environment is shifting. Beijing has been weighing tighter export controls on its top AI models and chips, a move that could restrict overseas access to precisely the kind of frontier systems ByteDance is now building. A 10-trillion-parameter model developed under those constraints would be a powerful, largely domestically-served system.
For the labs themselves, the calculation is increasingly about economics. Larger models are more capable but far more expensive to train and run, and the industry is split over whether raw scale or smarter architecture and post-training will determine the next winner.
The Bigger Picture
ByteDance's plan underscores a broader trend: the gap between the largest American and Chinese frontier models, once considered insurmountable, appears to be shrinking. If industry estimates of Mythos's size are accurate, a ByteDance model in the same ballpark would represent a meaningful closing of that distance.
It also raises the pressure on OpenAI, Anthropic, and Google to maintain their lead. With Chinese firms now targeting trillion-plus parameter systems and Western labs pursuing ever-larger architectures, the parameter race shows little sign of slowing — even as researchers debate how much the raw numbers truly matter.
Stay Ahead of the AI Race
The frontier-model landscape is moving faster than ever. Bookmark our homepage for the latest AI developments and in-depth coverage of every major release.
Read more AI news →
