DeepSeek Launching New Paper: Presenting New MHC Structures, List of Authors

January 2 news.DeepSeek A new article was publishedpaper, proposes a new structure called mHC. According to the presentation, the study aims to address the instability of traditional superconnection in large-scale model training while maintaining its significant performance gains。

The first three authors of the paper were Zhenda Xie, Yixuan Wei, Huanqi Cao. It's worth mentioning, DeepSeek founder and CEO Leung Man FungIt is also on the list of authors。

1AI HAS THE FOLLOWING SUMMARY:

• MORE RECENTLY, THE STUDY USING THE EXAMPLE OF HYPERCONNECTIVITY (HC) HAS EXPANDED THE UBIQUITOUS MODEL OF DISABILITY CONNECTIVITY ESTABLISHED OVER THE PAST DECADE BY EXPANDING THE WIDTH OF THE TRAFFIC AND THE DIVERSIFICATION OF CONNECTIONS. WHILE THERE HAS BEEN A SIGNIFICANT INCREASE IN PERFORMANCE, THIS DIVERSITY HAS FUNDAMENTALLY UNDERMINED THE CONSTANT EQUIVALENT MAPPING PROPERTIES INHERENT IN THE RESIDUAL CONNECTION, LEADING TO SERIOUS TRAINING INSTABILITY AND LIMITED SCALABILITY, AS WELL AS SIGNIFICANT MEMORY ACCESS COSTS。
• In order to address these challenges, we have proposed a current-form constraint superconnection (mHC), a common framework that can project the HC-discretion space onto a particular flow-form to restore the constant equivalent of mapping properties, while combining rigorous infrastructure optimization to ensure efficiency。
• Empirical experiments have shown that mHC is effective for large-scale training and provides practical performance improvements and scalability. We expect that the flexible and practical expansion of mHC as HC will contribute to a deeper understanding of the architecture design and provide promising directions based on the evolution of models。

Articles Link

• Hugging Face: https://huggingface.co/papers/2512.24880
• Arxiv: https://arxiv.org/abs/2512.24880