ByteDance Bets on a 10-Trillion-Parameter AI Leap to Challenge Anthropic

ByteDance Bets on a 10-Trillion-Parameter AI Leap to Challenge Anthropic
ByteDance is making its boldest move yet in the global AI race: a model so large it could put China’s most ambitious labs within striking distance of Anthropic’s frontier systems. The gamble is not simply about scale, but about whether a closed, independently built model can outrun faster-moving rivals.

The company is in the early stages of pre-training a system with as many as 10 trillion parameters, according to people familiar with the effort. That would be roughly three times the size of Moonshot’s Kimi K3, described as the largest Chinese model released so far, and could approach industry estimates for Anthropic’s Mythos 5, at about 8 trillion parameters.

The figure remains provisional. Pre-training commonly lasts three to six months, and ByteDance would settle the final size only later, before fine-tuning and a potential release. Parameter count signals a model’s capacity, but it does not guarantee superiority: training methods and data quality can be just as decisive.

The push comes as Chinese developers show sharper benchmark results, with models from Moonshot and Alibaba said to trail Anthropic’s Fable 5 only in certain areas. ByteDance, whose consumer model Doubao has 324 million monthly active users in China, has nonetheless kept much of its AI work closed rather than following domestic peers into open releases.

Over the past three years, ByteDance has expanded data centers, recruited researchers and built out Volcano Engine, its enterprise AI unit. Its Seed model team, led by former Google DeepMind scientist Wu Yonghui, now has about 2,000 members across China and overseas.

Crucially, Seed has pursued a more independent route for more than a year, avoiding the practice of distilling other labs’ models. That may have slowed progress, but founder Zhang Yiming has framed it as the only path to a breakthrough. In a recent internal meeting, he told the team to aim for “world-leading model capabilities” over the long term, rather than panic about near-term gaps.

Continue reading https://foxvector.com/stories/019fddd3-bff5-0618-7146-0f695da55673