ByteDance Training 10 Trillion-Parameter AI Model to Rival Anthropic's Mythos

ByteDance is training a massive artificial intelligence model with up to 10 trillion parameters, aiming to rival Anthropic's Mythos 5, which is estimated at 8 trillion parameters. The model is in its early pre-training phase, which is expected to last three to six months. This development comes amid intense US-China tech rivalry, marked by mutual accusations of AI distillation. ByteDance founder Zhang Yiming has urged his 2,000-member Seed team to focus on independent, long-term capabilities rather than short-term gains.

ByteDance 10 trillion-parameter model

• ▪ByteDance is in the early stages of training a mega AI model with up to 10 trillion parameters, which is currently in a pre-training phase expected to last three to six months
• ▪ByteDance's AI development team, Seed, is led by former Google DeepMind scientist Wu Yonghui and consists of approximately 2,000 members globally
• ▪ByteDance founder Zhang Yiming instructed the Seed development team in late July 2026 to focus on long-term world-leading capabilities and avoid using model distillation for short-term gains
• ▪ByteDance has invested aggressively in artificial intelligence by expanding its data center network, strengthening its Volcano Engine cloud unit, and pursuing custom AI chip development

US-China AI competition

• ▪The United States and China are locked in an intense technology rivalry characterized by mutual accusations of AI distillation, where models are trained on competitors' outputs
• ▪Chinese authorities have accused the United States of training artificial intelligence models on Chinese data and have threatened to implement countermeasures
• ▪The White House accused Chinese AI startup Moonshot in July 2026 of stealing Anthropic's Fable 5 technology to build its Kimi K3 model, an allegation that China denied

Anthropic Mythos comparison

• ▪Industry estimates place Anthropic's Mythos 5 model at approximately 8 trillion parameters and its Fable 5 model at around 5 trillion parameters
• ▪ByteDance's planned 10 trillion-parameter AI model would be more than three times the size of Moonshot's 2.8 trillion-parameter Kimi K3 model
• ▪Anthropic's Mythos 5 model is available only to approved organizations following a temporary suspension in June 2026 due to security concerns

AI parameter scale limits

• ▪Actual AI model performance depends heavily on factors beyond parameter size, including training data quality, model architecture, and optimization techniques
• ▪AI model parameters represent the numerical settings learned from data, but researchers caution that parameter count alone does not guarantee superior capability or efficiency

Chinese AI lab progress

• ▪ByteDance's consumer AI chatbot, Doubao, is the most popular in China with 324 million monthly active users, while its SeeDance model is highly advanced in video generation
• ▪Chinese AI models from Alibaba and Moonshot have demonstrated strong benchmark performance, lagging behind Anthropic's Fable 5 only in specific areas
• ▪Multiple Chinese AI laboratories are currently training models comparable in size to Anthropic's Fable 5, with ByteDance leading in scale ambition