ByteDance training 10 trillion-parameter AI model, FT reports

ByteDance is training an artificial intelligence model with as many as 10 trillion parameters, a scale that could approach Anthropic's most advanced Mythos system, the Financial Times reported on Friday, citing people with knowledge of the matter.reuters+1

The project, currently in a pre-training phase expected to last three to six months, would make ByteDance's model more than three times the size of Chinese startup Moonshot AI's Kimi K3, which has 2.8 trillion parameters. If completed at that scale, it would rank among the largest AI systems ever developed.thenews+1

Scale and Context

While Anthropic does not publicly disclose parameter counts for its models, industry estimates place its flagship Mythos 5 at roughly 8 trillion parameters and Fable 5 at about 5 trillion. ByteDance's reported ambition would put it at or above the frontier of current AI development.finimize+1

Parameters are numerical values a model learns during training, enabling it to recognize patterns and generate responses. However, researchers note that scale alone does not determine capability — training data quality, architecture, and optimization also matter.newsbytesapp+1

Before Kimi K3, China's largest models included Meituan's LongCat-2.0 and DeepSeek's V4-Pro, each with around 1.6 trillion parameters.timesnownews

Independent Development Over Distillation

The report arrives alongside a directive from ByteDance founder Zhang Yiming, who told staff to stop improving models by distilling the outputs of rival systems, instead favoring long-term independent capability over short-term gains. The order follows accusations from Anthropic that several Chinese AI firms used Claude at scale to train competing models.news.sbs+1

ByteDance's Seed research team, established in early 2023, now employs about 2,000 people across China and overseas. Wu Yonghui, who previously worked at Google Alphabet Inc. ↘ 3.58% and DeepMind, joined as the team's research head in February 2025.news.sbs

The Broader AI Race

The development underscores the intensifying competition between Chinese and American technology companies at the frontier of AI. In recent weeks, models from Moonshot and Alibaba ↘ 4.74% have received favorable reviews in comparative tests, trailing only Anthropic's Fable 5 in certain areas.news.sbs

Unlike many Chinese companies that favor open-source AI, most of ByteDance's models remain closed-source. Its Doubao chatbot is the most popular AI model in China with 324 million monthly active users, and its Seedance video generation model is considered among the world's most advanced.news.sbs

ByteDance did not respond to a request for comment, the Financial Times reported.news.sbs