ByteDance Trains 10 Trillion-Parameter AI Model to Rival Anthropic
The TikTok parent company is scaling its latest AI system to more than triple the size of other leading Chinese models.
ByteDance is reportedly training a massive new artificial intelligence model featuring approximately 10 trillion parameters. The move signals an aggressive attempt by the TikTok parent company to narrow the gap between Chinese AI capabilities and the current leaders in the United States.
The scale of the project is designed specifically to rival Anthropic's advanced "Mythos" system. According to reports from Ars Technica, Cybernews, and Techloy, the 10-trillion-parameter architecture would make it one of the largest AI systems ever constructed. To put the scale into perspective, the model is more than three times the size of Moonshot AI's Kimi K3, which possesses 2.8 trillion parameters.
The Race for Scale
ByteDance's push into ultra-large-scale modeling comes as the global AI race shifts toward increasing parameter counts to unlock higher-order reasoning and complex multimodal capabilities. For years, Chinese AI firms have been in a "catching up" phase, attempting to replicate the breakthroughs seen in U.S. frontier models from OpenAI and Anthropic. By investing in a model of this magnitude, ByteDance is attempting to move beyond imitation and potentially surpass Western competitors in raw computational power and reasoning depth.
Industry Implications
If successful, a 10-trillion-parameter model would represent a staggering investment in both high-end compute and curated data. Such a leap in scale could shift the balance of AI power toward Chinese firms, challenging the current dominance of U.S.-based models in specialized tasks and complex problem-solving. The sheer size of the model suggests that ByteDance is betting on the "scaling law"—the theory that increasing data and parameters consistently leads to emergent intelligence and better performance.
What to Watch
While the scale of the model is clear, the industry is now waiting to see how ByteDance manages the immense energy and hardware requirements necessary to sustain such a system. Observers will be looking for benchmarks that prove the 10-trillion-parameter count translates into actual performance gains over smaller, more efficient models. It remains to be seen when the model will move from its current training phase into a deployable product or a public API.