DeepSeek proposes a new AI training method

Chinese company DeepSeek introduced a new method in 2025 to save resources when training foundational AI models, 3dnews.ru reports. The method, called “multidimensional constraint with hyperconnections” (mHC), helps improve the cost efficiency of models. It enables performance on par with competing U.S. AI solutions, which typically have far greater compute resources.

DeepSeek’s paper notes that mHC was tested on models with 3B, 9B, and 27B parameters and resulted in a significant reduction in computational load. During model operation, the approach helps establish secure and fast connections using the existing hyperconnections method.

TEXT_2