DeepSeek Unveils New Architecture to Train Larger AI Models Cheaply
Chinese AI start-up DeepSeek kicks off 2026 by publishing a new technical paper that introduces a cost-effective deep learning architecture called Manifold-Constrained Hyper-Connections.
Chinese artificial intelligence start-up DeepSeek starts 2026 with a new technical paper co-authored by founder Liang Wenfeng that proposes a fundamental rethink of deep learning architecture. The new method, called Manifold-Constrained Hyper-Connections (mHC), aims to make foundational AI models more cost-effective as the company competes with better-funded US rivals.
This publication highlights the increasingly open and collaborative culture among Chinese AI companies, which continue to share a growing portion of their research publicly. Industry watchers closely follow DeepSeek's papers because they often provide early signals of the engineering choices that shape the start-up's next major model releases.
In the recent paper, a team of 19 DeepSeek researchers tests mHC on models with 3 billion, 9 billion, and 27 billion parameters. The empirical results confirm that mHC enables stable large-scale training and offers superior scalability compared to conventional hyper-connections, all without adding a significant computational burden.