ByteDance is pre-training a 10-trillion-parameter model to challenge frontier AI

ByteDance is pre-training an AI model with up to 10 trillion parameters, according to the Financial Times, citing three people with knowledge of the project. The scale would put the TikTok parent company within striking distance of the world's most capable AI systems and far ahead of any publicly known Chinese model.

The model would be roughly three times the size of Moonshot AI's Kimi K3 at 2.8 trillion parameters — currently China's largest known model. It also lands in the same ballpark as Anthropic's Mythos 5, which industry estimates place at around 8 trillion parameters. Anthropic has never publicly confirmed that number.

Pre-Training Already Underway

The model is currently in the pre-training phase, a process that typically takes three to six months at this scale. ByteDance has not publicly disclosed a name or a release schedule. The size alone signals that this is a long-term frontier bet, not an incremental product update.

One notable detail: ByteDance has avoided distillation — training on outputs from other companies' models — for more than a year. Distillation is a common shortcut for catching up quickly, but it constrains originality and creates intellectual property exposure. Avoiding it implies ByteDance is investing in full, original pre-training from scratch, which is significantly more expensive but produces more capable and defensible results.

CEO Acknowledges the Gap

The ambition of the training run stands in contrast to a candid internal admission. At a mid-year all-hands on August 6, 2026, ByteDance CEO Liang Rubo told staff that the company's large language models are currently falling behind leading overseas competitors. Rubo stressed a long-term commitment to in-house R&D over short-term gains from external model distillation.

ByteDance founder Zhang Yiming has reportedly told the 2,000-person Seed team — the division responsible for this model — to aim for world-leading capabilities over the long term. The company is increasing AI investment significantly in 2026, with a substantial share earmarked for semiconductor spending to support model training.

Not the Only One at This Scale

ByteDance isn't alone in pursuing trillion-parameter scale. Elon Musk's xAI is reportedly training Grok variants at 6 trillion and 10 trillion parameters on its Colossus 2 cluster. OpenAI and Google have not publicly disclosed parameter counts for their current systems, but competition at the multi-trillion parameter scale has become the quiet benchmark of the next generation of AI labs.

Parameter count isn't the only variable — data quality, training methods, and architecture matter just as much — but the signal from the Financial Times report is clear: ByteDance is no longer content to follow the frontier from a distance.