DeepSeek Proposes New Method to Reduce Model Training Costs

DeepSeek Launches Innovative Proposal in 2026

The Chinese artificial intelligence start-up, DeepSeek, kicked off 2026 with a new technical paper that reassesses the fundamental architecture used for training AI models. Co-authored by founder Liang Wenfeng, the study introduces an innovative approach to optimize model training more economically.

Proposed Method: Manifold-Constrained Hyper-Connections

The method, called Manifold-Constrained Hyper-Connections (mHC), is part of the strategy of the company based in Hangzhou, aiming to make its models more financially accessible. This initiative arises in a context where competition with American rivals, who have greater access to computational power, is intensifying.

Context and Implications

The proposed technological advancement seeks not only to reduce costs but also to expand the capabilities of AI models in a market that demands efficient and sustainable solutions. According to Liang Wenfeng, this approach is crucial for the evolution of the artificial intelligence investment sector.

Future Impact of Technology

The implementation of mHC could not only enhance DeepSeek's competitiveness but also influence the development of future AI architectures. The ability to train larger models in a less costly manner could democratize access to technology, benefiting smaller companies and startups.