DeepSeek Kicks Off the Year With New Paper Introducing the mHC Architecture
Want to read in a language you're more familiar with?
DeepSeek has unveiled a new research paper proposing mHC, a novel architecture designed to stabilize large-scale model training while preserving performance gains.
On January 1, 2026, DeepSeek released a new research paper introducing a novel architecture called mHC (manifold-constrained hyperconnection). The work aims to address the training instability of traditional hyperconnections (HC) in large-scale models, while retaining their significant performance benefits.
According to the paper, mHC projects the residual connection space of HC onto a specific manifold, restoring identity-mapping properties while incorporating rigorous infrastructure-level optimizations to ensure runtime efficiency. Empirical results show that mHC can effectively support large-scale training, delivering clear performance improvements alongside stronger scalability.
DeepSeek expects mHC, as a flexible and practical extension of hyperconnections, to deepen understanding of topological architecture design and point to promising directions for the evolution of foundation models.
The paper lists three co–first authors—Zhenda Xie, Yixuan Wei, and Huanqi Cao. Notably, DeepSeek founder and CEO Liang Wenfeng also appears among the authors.
