DeepSeek AI has released a new research paper that could shape its next generation of models, outlining a framework designed to make large-scale AI systems more efficient and scalable. Co-authored by founder Liang Wenfeng, the paper introduces Manifold-Constrained Hyper-Connections (mHC), a technique aimed at reducing the computational and energy demands of training advanced models while still supporting growth in model size and capability. Published on both arXiv and Hugging Face, the work is already fuelling expectations that DeepSeek may be preparing to unveil a successor to its R1 reasoning model, potentially around the Spring Festival.
The study continues a familiar pattern for the Hangzhou-based start-up, which has previously used academic publications as a signal ahead of major product launches. R1, which surprised the global AI community with its reasoning abilities, followed a similar trajectory. With mHC, DeepSeek appears to be laying the groundwork for what many observers believe could be an R2 model, reinforcing its reputation for unconventional approaches to AI development.
Authored by a team of 19 researchers, the paper reflects how Chinese AI labs are responding to ongoing chip export restrictions while still competing with leading US players such as OpenAI. Rather than relying purely on brute-force scaling, the research emphasizes architectural and infrastructure-level innovation. The authors describe testing the approach on models ranging from 3 billion to 27 billion parameters, paired with what they call “rigorous infrastructure optimization to ensure efficiency.”
The framework builds on earlier work around hyper-connections, including research from ByteDance, and seeks to better constrain how information flows through large neural networks. By structuring these connections more efficiently, the team argues that models can achieve stronger performance without a proportional increase in training cost or energy consumption—an increasingly important consideration as models grow larger and environmental scrutiny intensifies.
Beyond its technical contribution, the paper also highlights Liang Wenfeng’s continued hands-on role in shaping DeepSeek’s research agenda. Positioned as another example of the company’s non-traditional innovation strategy, the authors note that the technique holds promise “for the evolution of foundational models.” As global competition in AI intensifies, DeepSeek’s focus on efficiency-first scaling may become a defining feature of its efforts to remain competitive despite external constraints.
