ByteDance Trains 10 Trillion Parameter Model to Rival Anthropic

ByteDance is training an AI model with as many as 10 trillion parameters, a scale that would make it one of the largest ever built and a direct challenger to Anthropic's most advanced systems, according to a Financial Times report published Friday.

The TikTok owner is in the early pre-training stage of the model, which is estimated to be three times larger than Moonshot's Kimi K3, the biggest Chinese model released so far. Pre-training typically takes three to six months before fine-tuning begins, meaning the model is unlikely to launch before late 2026 or early 2027.

Anthropic does not publicly disclose parameter counts, but industry estimates place its most advanced Mythos 5 at roughly 8 trillion parameters and Fable 5 at around 5 trillion. ByteDance's 10-trillion-parameter target would surpass both on raw scale, though capability depends equally on data quality, architecture, and training methodology.

The effort is led by ByteDance's Seed team, a 2,000-person unit headed by former Google DeepMind scientist Wu Yonghui. The team includes core researchers, infrastructure engineers, and data labelling specialists spread across China and overseas offices.

ByteDance is taking a deliberately independent approach. Unlike some competitors that distill outputs from larger models to accelerate development, the company has avoided model distillation for over a year. Founder Zhang Yiming reinforced this stance in an internal meeting two weeks ago, telling the Seed team to target "world-leading model capabilities" in the long term without worrying about near-term benchmarks.

The company already commands China's largest AI consumer base. Doubao, its flagship chatbot, serves 324 million monthly active users, while its SeeDance model ranks among the most advanced globally in video generation. ByteDance has also built out its cloud unit, Volcano Engine, to sell enterprise AI solutions and has ambitions to develop custom AI chips.

What It Means for AI Builders

ByteDance's move is the clearest signal yet that the Chinese AI ecosystem is shifting from catching up to competing at the frontier. Multiple Chinese labs are now training models at Fable 5 scale, according to industry sources — and ByteDance is pushing furthest. For builders deploying AI agents and automation pipelines, this means the model landscape is about to get significantly more competitive. A new, independently developed frontier model from ByteDance would offer an alternative to the Anthropic-OpenAI duopoly, potentially driving down inference costs and expanding API access, particularly in Asian markets where ByteDance's infrastructure is deeply embedded.