ByteDance is Betting Big on a 10-Trillion-Parameter Model

ByteDance is making a bold move in the frontier-model race. According to the report, the company is training a model that could reach as many as 10 trillion parameters, putting it in the same broad class as the largest systems being built by leading US labs.

Key points

• The model is still in pre-training, a stage that usually lasts three to six months before any fine-tuning or release.
• Industry estimates put Anthropic’s most advanced Mythos 5 at around 8 trillion parameters, with Fable 5 around 5 trillion.
• Parameter count does not determine capability on its own, but it does set the model’s capacity and memory ceiling.
• ByteDance’s Seed team is following a more independent development strategy and is reportedly not relying on distillation from other labs’ models.
• The company has also been expanding its AI infrastructure, cloud offerings, and ambitions around custom chips.

Why it matters

This is another sign that the competition among top AI labs is shifting from product features to core training scale and infrastructure. Chinese companies are no longer just narrowing the gap in visible benchmark performance; they are now trying to build systems that can stand beside the biggest US models in raw size and ambition.

For ByteDance, the bet is strategic. A stronger base model could support its consumer AI products, enterprise cloud business, and broader AI ecosystem. It also reflects founder Zhang Yiming’s view that long-term leadership requires independent model development rather than shortcutting through distillation.

That said, bigger is not automatically better. Final performance will still depend on data quality, training strategy, and engineering execution. In other words, the 10-trillion-parameter target is a statement of ambition, not a guarantee of success.

Ars Technica AI