- ByteDance wants to get ahead of US competitors in AI.
- The company is preparing a 10 trillion-parameter model.
- Unlike other players in the Chinese market, it prioritizes developing closed AI models.
China’s ByteDance, TikTok’s parent company, has begun training a large AI model with potentially up to 10 trillion parameters, aiming to narrow the gap with leading US developers, FT reports.
According to three sources familiar with the matter, the model is currently in the pre-training stage, and its final size will be determined later. Such a scale could make it one of the largest AI models in the world and roughly triple the parameter count of China’s Kimi K3.
ByteDance Sets Its Sights on Anthropic-Scale
Pre-training a model typically takes three to six months. After that, ByteDance plans to further fine-tune the system and release it, if the development process stays on track.
For comparison, industry experts estimate Anthropic’s most powerful models could have around 8 trillion parameters. At the same time, the company does not officially disclose the parameter counts of its models.
The number of parameters determines a model’s fundamental capacity to store and process information, but it does not, by itself, guarantee higher quality. AI capabilities are also influenced by:
- The quality and volume of training data
- Training methods
- Model architecture
- Approaches to further fine-tuning and optimization
Therefore, the potential 10 trillion parameters primarily signal the scale of ByteDance’s infrastructure and ambitions, rather than automatically indicating an advantage over competitors.
Chinese developers have been ramping up competition with US AI labs in recent months. In particular, Moonshot and Alibaba models have posted strong results across a range of benchmarks, trailing only Anthropic’s cutting-edge systems in some categories.
At the same time, several Chinese labs, according to sources, are also working on models at a scale of around 5 trillion parameters. ByteDance may currently be the company with the most ambitious plan to increase the parameter count.
ByteDance Bets on In-House Technology
Despite the scale of the project, ByteDance discloses far less about its AI work than some Chinese rivals. The company primarily develops closed models.
Its video-generative model Seedance is already among the most powerful systems in the world in its segment, while the consumer model Doubao is the most popular in China and has around 324 million monthly active users.
Recall that Seedance 2.0, released in February 2026, sparked a wave of interest and intensified talk of a “second DeepSeek moment.”
Over the past three years, ByteDance has been ramping up investment in AI, expanding its data center network and recruiting researchers. The company is also strengthening its cloud unit, Volcano Engine, which sells AI-based solutions to enterprise clients, and is working on its own AI chips.
Model development is handled by the Seed team led by former Google DeepMind researcher Wu Yonghui. According to sources, it has around 2,000 employees in China and abroad, including researchers, infrastructure engineers, data labeling specialists, and translators.
At the same time, ByteDance is betting on independent model development and does not use distillation of other systems. This approach involves training a smaller model on the knowledge and answers of a larger “teacher” model and is widely used to build more compact AI.
According to a source, ByteDance management believes that independent development is what could enable the company to build a system capable of outpacing competitors.
ByteDance founder Zhang Yiming reiterated this position at an internal meeting about two weeks ago. He urged the Seed team, over the long term, to aim for “world-leading model capabilities” and not to worry too much about potentially lagging behind competitors at the current stage.
Previously, we wrote that ByteDance plans to spend up to $170 billion on AI infrastructure.
Сообщение FT: TikTok’s Creator Started Training One of the World’s Largest AI Models появились сначала на INCRYPTED.