Back
ByteDance reportedly develops AI model that may reach 10 trillion parameters

ByteDance reportedly develops AI model that may reach 10 trillion parameters

The early-stage project would be about three times larger than Moonshot's Kimi K3 and extends ByteDance's strategy of accepting short-term LLM lag while pursuing independent development.

Fact Check
The Benzinga report (citing the Financial Times) and Reuters-sourced coverage carried by WTVB and The Standard all corroborate the core claim: ByteDance is reportedly training an AI model with up to 10 trillion parameters, roughly three times the size of Moonshot's Kimi K3 (2.8 trillion), in an early pre-training phase (3-6 months) with no finalized size or release date. This matches the claim's assertion of an early-stage project about three times larger than Kimi K3 and its independent-development framing. All coverage traces to the same primary FT report; the claim is accurately characterized as reported/early-stage rather than confirmed product, consistent with the sources.
    Reference: 123
Summary

ByteDance is reportedly pre-training an AI model that may reach 10 trillion parameters, a scale that would make it about three times larger than Moonshot AI's Kimi K3, China's largest released AI model, and potentially comparable to Anthropic's most advanced Mythos systems. The project remains in an early stage, with pre-training typically lasting three to six months before ByteDance decides the model's final size, potential fine-tuning and any release. The effort reinforces ByteDance's stated preference for independent development: founder Zhang Yiming has urged the AI team to pursue a world-class model rather than short-term competition, while CEO Liang Rubo has said the company's core LLMs are falling further behind top overseas rivals. ByteDance has nonetheless built strong positions in applications, with Doubao leading China's consumer AI market at 324 million monthly active users, SeeDance viewed as one of the world's most advanced video-generation models, and tighter ties among Doubao, Feishu and Volcano Engine helping enterprise sales.

Terms & Concepts
  • Parameters: Internal numerical weights in an AI model that are adjusted during training and help determine how much information the model can represent.
  • Pre-training: The initial training phase in which a model learns broad patterns from large datasets before later tuning for specific tasks.
  • LLM: Large language model, an AI system trained on vast amounts of text to understand and generate language.