ByteDance Is Building a 10-Trillion-Parameter AI Model

The Chinese corporation ByteDance, owner of TikTok, has begun training an AI model that could reach 10 trillion parameters in size. That is three times larger than Moonshot's Kimi K3 — the biggest Chinese model released to date. Three sources familiar with the matter have described the company's plans.

The model is in the early stages of pre-training — a phase that typically takes three to six months, followed by fine-tuning and, if all goes according to plan, release. The exact size won't be settled until later. For comparison: Anthropic does not disclose the parameters of its systems, but by industry estimates its flagship Mythos 5 comes in at around 8 trillion parameters, and Fable 5 at roughly 5 trillion.

It's worth remembering that the parameter count only defines a model's fundamental capacity — its ability to store information. Real-world capabilities also depend on the quality of the data and the training methods.

The Race for the Top

Launching a project on this scale reflects the ambition of Chinese labs not merely to catch up with the leading American companies but to surpass them. In recent weeks, models from Moonshot and Alibaba have posted strong benchmark results, trailing only Fable 5 in certain areas. The most advanced model, Mythos 5, is available only to approved organizations — following a temporary ban in June over safety concerns.

According to insiders, several Chinese labs are training Fable 5–class models at once, but it is ByteDance that has set its sights on the largest of all.

The Quiet Giant

Unlike many of its compatriots, ByteDance keeps a low profile: its models are largely closed. Even so, its SeeDance video generator ranks among the best in the world, and its flagship consumer service, Doubao, is the most popular in China, with 324 million monthly active users. Over the past three years the company has invested in AI more aggressively than any other Chinese tech giant, building data centers, hiring researchers, and growing its Volcano Engine cloud division. It also has plans for its own AI chips.

The Seed division, led by former Google DeepMind scientist Wu Yonghui, has around 2,000 employees. Notably, for more than a year ByteDance has made a point of refusing to “distill” other companies' models — training a compact network to imitate a larger one. Some believe it is precisely this go-it-alone approach that has slowed its development compared with competitors.

The company's founder, Zhang Yiming, is convinced, however, that only independent development can produce a model that outdoes its rivals. At an internal meeting two weeks earlier, he urged the Seed team to aim for “global leadership” over the long term and not to fret over falling behind in the short run.

Source: Ars Technica