China’s AI developers deepen push into massive foundation models, backed by lower costs and investor momentum despite US restrictions
Chinese artificial intelligence developers are accelerating their push into massive foundation models with more than a trillion parameters, just as Washington moves to block foreign access to leading US software through unprecedented export controls.
Parameters serve as a primary measure of an AI’s capabilities. Chinese companies have been looking to narrow the gap with leading US rivals like OpenAI and Anthropic, which continue to aggressively expand the size of their top models.
The depth of the US-China tech divide was underscored late last week, when Anthropic suspended global access to Mythos and Fable following export controls imposed by the Trump administration. Both top-tier models are estimated to have trillions of parameters.
Chinese developers were rapidly moving away from billion-parameter general models popular in 2023 and 2024 to trillion-parameter architectures featuring million-token contexts and full adaptation to domestic chip stacks, according to a report by Donghai Securities.
In late April, Chinese AI champion DeepSeek launched its first trillion-parameter model, V4.
Other Chinese tech giants, including Xiaomi and Alibaba Group Holding, have also rolled out trillion-parameter models in recent months. Alibaba was among the first to cross the threshold with Qwen-3-Max-Preview in September. Alibaba owns the South China Morning Post.