ByteDance is training a 10-trillion-parameter model to chase the frontier

ByteDance is going big, in the most literal sense. The owner of TikTok is training an AI model with around 10 trillion parameters, a scale that would put it among the largest ever built, and it is not hiding what it wants that scale for.

According to the Financial Times, the model is meant to rival Anthropic’s Mythos, one of the frontier systems that Chinese developers have so far struggled to match.

The size is itself the statement. At roughly 10 trillion parameters, the model would be more than three times as large as Moonshot’s Kimi K3, which sits among the biggest Chinese models today at about 2.8 trillion.

That gap is not a rounding difference; it is the kind of leap that reorders where a model ranks against the field. It is still early, though: the project is said to be in the early pre-training phase, a stage that can take three to six months, with fine-tuning and further work to come before anything is released.

Parameter count is not everything, of course. Bigger models are not automatically better, and the industry has learned that data quality, training technique and efficiency often matter as much as raw scale.

Even so, committing the compute to train a model this size is a declaration in its own right, a signal that ByteDance wants to compete at the very top rather than ship a capable also-ran.

It has some ground to stand on. Its Doubao assistant is already one of China’s most used AI products, which gives the company both a distribution channel and a reason to want a frontier model of its own.

Behind the effort sits a pointed instruction, too: founder Zhang Yiming reportedly told staff to avoid leaning on AI distillation for short-term gains, pushing them to build genuine capability rather than copy it.

That warning lands amid a live controversy, since US labs have accused Chinese firms of distilling their models.

Moonshot, for one, is alleged to have leaned on Anthropic’s work for Kimi K3, claims China disputes and answers with its own. The fight has spilled into tooling as well, and concerns about Chinese coding tools and Claude Code show how tangled the two ecosystems have become even as they compete.

ByteDance is far from the only Chinese firm scaling up. Alibaba has been pushing Qwen toward the top ranks and claims its latest is the world’s number-two open-weight model, while the sector as a whole races to close the gap.

The compute those ambitions demand is staggering, and Moonshot reportedly used 20,000 Nvidia chips to train Kimi K3, which hints at what a far larger model will consume.

Access is becoming a strategic lever, too. China has weighed curbing overseas access to its best models, a sign that these systems are increasingly treated as national assets rather than mere products.

Hanging over all of it is the chip question, because US export controls have limited China’s access to the most advanced accelerators, so training a model this large tests how far Chinese firms can push with the hardware they can actually get.

ByteDance does bring advantages others lack. Its enormous consumer reach through TikTok and Doubao supplies both training data and a ready place to deploy a frontier model, turning research spending into products almost immediately.

There is a strategic subtext beyond the balance sheet as well, since a Chinese company matching a top Western model would be a symbolic milestone in a rivalry both governments increasingly frame in national terms.

For now, the model remains a work in progress, and much can still change between an early pre-training run and a finished system.

Yet the sheer size of the target shows how determined China’s giants are to stand alongside the West’s frontier labs rather than trail behind them.