ByteDance is developing an artificial intelligence model that could rival the scale of Anthropic’s most advanced systems, highlighting China’s growing ambition to compete directly with the world’s leading AI laboratories as the global race for increasingly powerful foundation models accelerates.
According to three people familiar with the matter, the Chinese technology company has begun the early stages of training a model containing as many as 10 trillion parameters, making it potentially three times larger than Moonshot’s Kimi K3, currently the largest Chinese AI model released to date. If completed, the model would also exceed industry estimates for the size of Anthropic’s flagship Mythos 5 system, which is believed to contain around 8 trillion parameters.
The model is currently undergoing pre-training, a phase of development that typically lasts between three and six months before fine-tuning begins. One person familiar with the project said the final number of parameters would only be determined later in the development process, depending on training outcomes.
Although Anthropic does not publicly disclose the size of its models, industry estimates suggest that its most advanced Mythos 5 model contains approximately 8 trillion parameters, while Fable 5 is believed to have around 5 trillion. While parameter count determines the theoretical capacity of an AI model to store and process information, experts note that overall performance also depends heavily on factors such as data quality, training techniques and system architecture.
ByteDance’s latest effort underscores the determination of Chinese AI developers not merely to narrow the technological gap with the United States but ultimately to surpass leading American companies in frontier artificial intelligence. Recent releases from Chinese firms, including Moonshot and Alibaba, have demonstrated increasingly competitive benchmark performance, trailing only Anthropic’s Fable 5 in certain areas. Anthropic’s most advanced model, Mythos 5, remains available only to approved organisations after the company temporarily restricted broader access in June because of security concerns.
Industry insiders say several Chinese AI laboratories are currently training models comparable in scale to Fable 5. However, ByteDance is regarded as the most ambitious among them, pursuing what could become one of the largest AI models developed anywhere in the world.
Despite maintaining a relatively low public profile in artificial intelligence compared with many Chinese competitors, ByteDance has steadily expanded its capabilities. Most of its models remain closed rather than open-source, distinguishing its approach from several domestic rivals. Its SeeDance model is considered among the world’s leading video-generation systems, while its flagship consumer chatbot, Doubao, has become China’s most widely used AI application, attracting 324 million monthly active users.
Over the past three years, ByteDance has invested more aggressively in AI than any other major Chinese technology company, significantly expanding its data centre infrastructure and recruiting large numbers of researchers. The company has also strengthened its cloud computing division, Volcano Engine, which provides AI services to enterprise customers, while pursuing plans to develop proprietary AI chips.
The company’s model development is led by Seed, a research organisation headed by former Google DeepMind scientist Wu Yonghui. The team employs approximately 2,000 people across China and overseas, including AI researchers, infrastructure engineers, data labelling specialists and translators.
According to one person familiar with the project, Seed has adopted an independent research strategy that avoids the practice of “model distillation”, whereby smaller models are trained by learning from the outputs of larger existing systems. That approach has been in place for more than a year and is believed by some observers to have contributed to ByteDance progressing more slowly than some competitors in releasing new frontier models.
Model distillation has become a widely used technique for compressing large AI systems into smaller and faster versions while preserving much of their capability. ByteDance’s leadership, however, believes that genuine technological leadership can only be achieved through independently developed models rather than by relying on existing systems created elsewhere.
Founder Zhang Yiming recently reinforced that strategy during an internal meeting with the Seed team, according to one person familiar with the discussion. He reportedly urged researchers to pursue “world-leading model capabilities” over the long term and not become overly concerned about temporarily lagging behind competitors as the company focused on building stronger foundational technologies.
Chinese media outlet Latepost and The Information previously reported Zhang’s comments from the internal meeting. ByteDance did not respond to a request for comment.
The company’s latest initiative reflects the increasingly intense competition shaping the global AI industry, with Chinese developers investing heavily in large-scale models as they seek to challenge the technological dominance of leading US companies and establish themselves at the forefront of next-generation artificial intelligence.
