In 2025, despite DeepSeek not rolling out its new large - scale model as planned, the team persisted in refining the DeepSeek V3.X series of large - scale models and unveiled a number of novel technologies. On January 1st, 2026, DeepSeek released a new paper, introducing the 'Manifold - Constrained Hyper - Connection' (mHC) framework. This framework is specifically designed to tackle the instability problems that often plague large - scale model training, all while preserving the performance enhancements.

The mHC framework achieves training stability by employing mathematical constraints. At the same time, it ensures efficiency through system - level optimizations. Experimental results demonstrate that the mHC framework brings about substantial performance improvements. Moreover, it offers better scalability, with just a 6.7% increase in training time. This groundbreaking innovation sheds new light on the architectural design of next - generation foundational models and has the potential to propel further progress in the global AI technology landscape.