Blab AI
ByteDance Pursues 10-Trillion-Parameter AI Model to Match U.S. Frontier Systems
ByteDance, the Chinese company behind TikTok, is reportedly training a new large‑language model that could contain up to 10 trillion parameters, according to a Financial Times report published on 10 August 2026. The model is described as a mixture‑of‑experts architecture that is being pre‑trained on roughly 30,000 GPUs, a scale that would put it in the same class as Anthropic’s Mythos 5, which is estimated to have about 8 trillion parameters.
The development effort is part of ByteDance’s broader AI strategy, which includes building new data‑center infrastructure and expanding its research teams. The company’s internal AI unit, known as the Seed team, focuses on foundational models, multimodal capabilities, and next‑generation AI interactions. The 10‑trillion‑parameter project is said to be in the early stages of training, and it is not yet a commercial product.
Chinese AI firms have been rapidly expanding the size of their models in recent months. Moonshot AI released Kimi K3 in July 2026, a 2.8‑trillion‑parameter model that includes native vision capabilities and a 1‑million‑token context window. Alibaba’s Qwen 3.8‑Max, released in June, contains 2.4 trillion parameters and performed well in the Arena.ai benchmark suite, ranking fifth in Text Arena and second in Vision Arena.
The new ByteDance model would surpass both Moonshot and Alibaba in parameter count and would bring the company closer to the scale of U.S. frontier models. The Financial Times noted that the development of such a model aligns with a broader trend of Chinese firms catching up to U.S. competitors in large‑language‑model research.
The U.S. AI community has expressed concern about the rapid growth of large models in China. In February 2026, Anthropic publicly accused several Chinese labs—including DeepSeek, Moonshot, and MiniMax—of conducting “industrial‑scale campaigns” to illicitly extract Claude model capabilities through a process known as distillation. Anthropic’s report described the alleged activity as a coordinated effort to replicate the performance of its models without direct access to the underlying training data.
ByteDance has stated that it does not use distillation techniques to accelerate its model development. According to a Chinese state‑backed newspaper, The Paper, ByteDance founder Zhang Yiming has emphasized a long‑term approach to AI research, discouraging the use of other companies’ outputs to achieve short‑term gains. The company reportedly enforces internal restrictions on the use of open‑source models for distillation and has implemented additional controls such as API testing.
The development of a 10‑trillion‑parameter model raises questions about the competitive dynamics between U.S. and Chinese AI firms. While the U.S. currently holds a lead in frontier model size, Chinese companies are closing the gap through large‑scale training projects and the release of open‑weight models. The race has attracted attention from policymakers and researchers concerned about the potential security and ethical implications of rapidly scaling AI systems.
ByteDance’s move also reflects the company’s broader ambitions beyond social media. Its 2026 AI roadmap includes goals such as achieving world‑model state‑of‑the‑art performance, improving coding data loops, defending its video‑editing lead, and monetizing its Doubao platform, which serves a large user base.
The 10‑trillion‑parameter project is still in the pre‑training phase, and no release date has been announced. The model’s eventual capabilities, deployment plans, and potential commercial applications remain unclear. Analysts will be watching for updates on the model’s progress, any public benchmarks, and how the project fits into ByteDance’s overall AI strategy.
In summary, ByteDance is advancing a large‑parameter AI model that could rival U.S. frontier systems, following the release of similarly sized models by other Chinese firms. The project is part of a broader trend of rapid scaling in China’s AI sector, which has prompted concerns about competitive practices and the use of distillation. The outcome of this effort will influence the balance of power in the global AI landscape and may shape future regulatory and security discussions.
The development effort is part of ByteDance’s broader AI strategy, which includes building new data‑center infrastructure and expanding its research teams. The company’s internal AI unit, known as the Seed team, focuses on foundational models, multimodal capabilities, and next‑generation AI interactions. The 10‑trillion‑parameter project is said to be in the early stages of training, and it is not yet a commercial product.
Chinese AI firms have been rapidly expanding the size of their models in recent months. Moonshot AI released Kimi K3 in July 2026, a 2.8‑trillion‑parameter model that includes native vision capabilities and a 1‑million‑token context window. Alibaba’s Qwen 3.8‑Max, released in June, contains 2.4 trillion parameters and performed well in the Arena.ai benchmark suite, ranking fifth in Text Arena and second in Vision Arena.
The new ByteDance model would surpass both Moonshot and Alibaba in parameter count and would bring the company closer to the scale of U.S. frontier models. The Financial Times noted that the development of such a model aligns with a broader trend of Chinese firms catching up to U.S. competitors in large‑language‑model research.
The U.S. AI community has expressed concern about the rapid growth of large models in China. In February 2026, Anthropic publicly accused several Chinese labs—including DeepSeek, Moonshot, and MiniMax—of conducting “industrial‑scale campaigns” to illicitly extract Claude model capabilities through a process known as distillation. Anthropic’s report described the alleged activity as a coordinated effort to replicate the performance of its models without direct access to the underlying training data.
ByteDance has stated that it does not use distillation techniques to accelerate its model development. According to a Chinese state‑backed newspaper, The Paper, ByteDance founder Zhang Yiming has emphasized a long‑term approach to AI research, discouraging the use of other companies’ outputs to achieve short‑term gains. The company reportedly enforces internal restrictions on the use of open‑source models for distillation and has implemented additional controls such as API testing.
The development of a 10‑trillion‑parameter model raises questions about the competitive dynamics between U.S. and Chinese AI firms. While the U.S. currently holds a lead in frontier model size, Chinese companies are closing the gap through large‑scale training projects and the release of open‑weight models. The race has attracted attention from policymakers and researchers concerned about the potential security and ethical implications of rapidly scaling AI systems.
ByteDance’s move also reflects the company’s broader ambitions beyond social media. Its 2026 AI roadmap includes goals such as achieving world‑model state‑of‑the‑art performance, improving coding data loops, defending its video‑editing lead, and monetizing its Doubao platform, which serves a large user base.
The 10‑trillion‑parameter project is still in the pre‑training phase, and no release date has been announced. The model’s eventual capabilities, deployment plans, and potential commercial applications remain unclear. Analysts will be watching for updates on the model’s progress, any public benchmarks, and how the project fits into ByteDance’s overall AI strategy.
In summary, ByteDance is advancing a large‑parameter AI model that could rival U.S. frontier systems, following the release of similarly sized models by other Chinese firms. The project is part of a broader trend of rapid scaling in China’s AI sector, which has prompted concerns about competitive practices and the use of distillation. The outcome of this effort will influence the balance of power in the global AI landscape and may shape future regulatory and security discussions.
