ByteDance is making its biggest visible bet yet on frontier AI: a model so large it is designed to put China’s most ambitious labs in direct competition with Anthropic. The gamble is not simply on scale, but on whether patience and self-reliance can beat a faster-moving pack.

The effort is still early. ByteDance is reportedly pre-training a system that could reach 10 trillion parameters — roughly three times the size of Moonshot’s Kimi K3, the largest Chinese model released so far. Pre-training alone can take three to six months, and the final size has not been fixed.¹

If it reaches that target, the model would exceed industry estimates for Anthropic’s Mythos 5, at about 8 trillion parameters. But the comparison comes with a caveat: Anthropic does not disclose model sizes, and parameter counts are only one measure of performance; data quality and training methods can matter just as much.¹

The chronology matters. Over the past three years, ByteDance has built data centers, expanded its Volcano Engine cloud business, recruited researchers and pursued custom chips. Its Seed model team now has about 2,000 people across China and overseas, while Doubao has become China’s most popular consumer-facing AI model, with 324 million monthly active users.¹

Yet ByteDance’s approach differs sharply from rivals that have moved quickly by distilling other models. Seed has spent more than a year pursuing independent development — a choice some insiders believe slowed progress, but one management sees as necessary for a genuine leap ahead. Founder Zhang Yiming recently told the team to pursue “world-leading model capabilities” over the long term and not become overly anxious about trailing in the near term.¹

That leaves ByteDance with a high-stakes test: a massive model may signal intent, but it will take more than raw parameter count to prove China can close the gap at AI’s frontier.