Chinese AI laboratory DeepSeek published a research paper on 1st January 2026 introducing Manifold-Constrained Hyper-Connections (mHC), a novel architecture designed to improve AI model performance whilst minimising training costs. The paper, co-authored by DeepSeek founder and CEO Liang Wenfeng, has been cited as a potential game changer in developing AI models.
The mHC architecture marks an improvement to conventional residual networks (ResNet), a fundamental mechanism underlying large language models, showcasing the Chinese AI start-up's continuous efforts to train powerful models with limited computing resources. In internal tests, DeepSeek determined that mHC incurs a hardware overhead of only 6.27%, whilst a team of 19 DeepSeek researchers tested mHC on models with 3 billion, 9 billion and 27 billion parameters and found it scaled without adding significant computational burden. According to DeepSeek, the mHC-powered large language models performed better across eight different AI benchmarks. Florian Brand, a PhD student at Germany's Trier University and an expert on China's AI ecosystem, said DeepSeek's papers often acted as an early signal of the technical direction behind its next generation of models, with industry expectations running high that DeepSeek could release its next major model in the run-up to the Spring Festival holiday in mid-February.
The Hangzhou-based start-up stunned the industry with the R1 reasoning model a year ago, developed at a fraction of the cost of its Silicon Valley rivals. The new announcement suggests that DeepSeek might have once again bypassed compute bottlenecks and unlocked leaps in intelligence.
Sources:
1. https://siliconangle.com/2026/01/01/deepseek-develops-mhc-ai-architecture-boost-model-performance/
2. https://dnyuz.com/2026/01/02/chinas-deepseek-kicked-off-2026-with-a-new-ai-training-method-that-analysts-say-is-a-breakthrough-for-scaling/
3. https://finance.yahoo.com/news/chinas-deepseek-kicked-off-2026-071041508.html
4. https://www.scmp.com/tech/tech-trends/article/3338535/deepseek-proposes-shift-ai-model-development-mhc-architecture-upgrade-resnet
5. https://sg.news.yahoo.com/deepseek-kicks-off-2026-paper-093000536.html
6. https://www.freemalaysiatoday.com/category/business/2026/01/02/deepseek-touts-new-training-method-as-china-pushes-ai-efficiency
