-3.png?w=1024)
ByteDance is reportedly training an AI model that could reach 10 trillion parameters, signaling a major escalation in the TikTok owner’s push into frontier artificial intelligence.
The model remains in pre-training and its final size has not been determined, according to the Financial Times. If it reaches the reported scale, it would rank among the largest AI models ever developed.
If it reaches the reported scale, the system would be more than three times larger than Moonshot AI’s Kimi K3, currently the largest publicly disclosed Chinese model at 2.8 trillion parameters. Industry estimates place Anthropic’s advanced Mythos 5 model at roughly 8 trillion parameters, though Anthropic does not publish official parameter counts.
ByteDance is betting on its own technology
The project is reportedly being led by Seed, ByteDance’s AI research team headed by former Google DeepMind scientist Wu Yonghui. The team has about 2,000 members across China and overseas.
One notable part of the strategy is ByteDance’s reported decision to avoid model distillation — the practice of training a model using outputs from another company’s AI system. According to the Financial Times, ByteDance management believes independent development is the only path to building a model capable of surpassing rivals.
Founder Zhang Yiming reportedly reinforced that approach in a recent internal meeting, telling the Seed team to pursue “world-leading model capabilities” over the long term, according to the Financial Times.
China’s AI race is accelerating
The reported effort comes as Chinese AI companies push aggressively to close the gap with leading U.S. labs such as Anthropic and OpenAI. Recent models from Moonshot and Alibaba have posted strong benchmark results, and several Chinese companies are reportedly working on systems around the 5-trillion-parameter scale.
ByteDance has kept a lower profile than many domestic rivals because most of its AI models are closed rather than open-weight. Its Doubao chatbot has about 324 million monthly active users in China, while its SeeDance video-generation model is considered one of the strongest in its category.
The company has also spent heavily on AI infrastructure, expanding data centers, hiring researchers, growing its enterprise AI business through Volcano Engine, and pursuing custom AI chip development.
What this could mean
The reported project matters less because of the raw number and more because of what it signals. Training a model of this size would require enormous computing resources, specialized infrastructure, and a long-term commitment to frontier AI research.
Parameter count alone does not determine how capable an AI model will be. Performance also depends on training data, architecture, optimization methods, and how efficiently the model uses its parameters. A 10-trillion-parameter system could still underperform a smaller but better-trained model.
If ByteDance succeeds, the model could give the company a stronger foundation for products across consumer AI, video generation, and enterprise services. More broadly, it would reinforce a trend already reshaping the industry: Chinese technology companies are increasingly willing to spend at frontier scale rather than simply follow the largest U.S. labs.
Other News: Hugging Face CEO Clément Delangue said China is already leading in open-weight AI and could gain a broader frontier-model advantage as Chinese developers build more openly on one another’s work.