Chinese AI startup DeepSeek has introduced a new way to train large AI models more efficiently. The approach, called Manifold‑Constrained Hyper‑Connections (mHC), helps models learn faster while using less computing power and energy. This is especially important as China faces limits on buying the latest AI chips from abroad due to US export restrictions.
The research paper, co‑authored by founder Liang Wenfeng and 18 other researchers, tested mHC on AI systems ranging from 3 to 27 billion parameters. The method stabilises training and avoids excessive computing costs, making it easier to build very large AI models without huge energy bills.
DeepSeek has a history of surprising the AI industry. Its 2025 R1 reasoning model was developed at a much lower cost than similar models from US companies. Experts now expect the next model, R2, to launch around China’s Spring Festival in February. The new mHC training method is expected to power this model, making it faster and more efficient.
China’s AI firms continue to face challenges due to limited access to advanced semiconductors. This has pushed companies like DeepSeek to create innovative, resource‑saving techniques to stay competitive globally.
Analysts suggest that R2 could make a significant impact internationally, even as companies like Google and OpenAI release high‑performance models. China’s lower-cost, efficient AI models are already gaining recognition in global rankings, showing the country’s growing technical capabilities.
DeepSeek has shared its research on open platforms like arXiv and Hugging Face, reflecting a trend of more openness and collaboration among Chinese AI developers.
The new method could set a benchmark for energy-efficient, large-scale AI training, helping China expand its AI capabilities despite hardware limitations.
Also Read: Aurobindo Pharma buys Khandelwal labs’ drugs for ₹325 cr