Chinese AI startup DeepSeek has released a research paper presenting a new approach to building scalable and energy-efficient AI systems. The framework, named Manifold-Constrained Hyper-Connections, is designed to reduce computational demands while improving training stability for large models.
The paper, co-authored by 19 researchers and led by founder Liang Wenfeng, was published on arXiv and Hugging Face. Tests were conducted on models ranging from 3 billion to 27 billion parameters, leveraging techniques inspired by ByteDance’s 2024 research into hyper-connection architectures. The authors highlight rigorous infrastructure optimization as a key component of the framework.
DeepSeek previously gained attention for its R1 reasoning model, developed at lower cost than many Silicon Valley counterparts. The company is expected to release its next flagship system, widely referred to as R2, around the Chinese Spring Festival in February.
The development underscores how Chinese AI firms are pursuing alternative architectures to maintain competitiveness amid restrictions on access to advanced US-made semiconductors. The company said the framework “holds promise for the evolution of foundational models” and reflects ongoing efforts to overcome resource constraints while advancing large-scale AI capabilities.
