DeepSeek is training a 2T-parameter model and plans to build an 8T-parameter one (twitter.com)

DeepSeek has announced its ambitious plan to train a groundbreaking 2-trillion (2T) parameter AI model, with intentions to scale up to an 8-trillion (8T) parameter version in the future. CEO Liang Wenfeng emphasized to investors that a crucial focus moving forward will be the employment of more domestic chips for AI training, marking a significant shift toward reliance on local technology. This comes as Huawei is expected to start delivering these domestic chips by the fourth quarter, providing a necessary infrastructure boost for DeepSeek's expansive AI endeavors. The commitment to expanding parameter sizes reflects a growing trend in the AI/ML community aimed at enhancing model capabilities and performance. Larger models generally lead to improved natural language processing, better understanding of complex data patterns, and increased overall efficiency. This development not only underscores DeepSeek's strategic alignment with national technological initiatives but also positions it to compete more effectively on the global stage as the demand for powerful, locally-sourced AI solutions intensifies.