On January 1st, 2026, DeepSeek published a groundbreaking paper introducing a revolutionary new architecture called mHC (Manifold-Constrained Hyper-Connection). This innovative architecture is designed to overcome the instability challenges commonly encountered with traditional hyper-connections during large-scale model training, all while delivering substantial performance improvements. The lead authors of this significant work are ZhenDa Xie, YiXuan Wei, and HuanQi Cao. Additionally, Liang Wenfeng, the visionary founder and CEO of DeepSeek, is also credited as one of the contributing authors.
