DeepSeek Touts New Training Method as China Pushes AI Efficiency
DeepSeek is a company that outlines a more effective approach to AI development and is part of the Chinese AI industry’s Nvidia Corp. published an article showing its effort to compete with companies like OpenAI despite the lack of free access to its chips.
The document, co-authored by founder Liang Wenfeng, introduces a framework called Manifold Constrained Hyperlinks. According to the authors, it is designed to increase scalability while reducing the computational and energy demands of training advanced AI systems.
Such publications from DeepSeek have heralded the release of major models in the past. The Hangzhou-based startup stunned the industry a year ago with its R1 reasoning model, developed at a fraction of the cost of its Silicon Valley rivals. DeepSeek has since released several smaller platforms, but expectations are rising for its next flagship system, commonly referred to as R2, expected around the Spring Festival in February.
Chinese startups continue to operate under significant restrictions as the United States blocks access to the most advanced semiconductors needed to develop and operate artificial intelligence. These constraints have forced researchers to pursue unconventional methods and architectures.
What Does Bloomberg Intelligence Say?
DeepSeek’s R2 model, which will be released in the next few months, has the potential to disrupt the global AI industry again, despite Google’s recent gains. Google’s Gemini 3 surpassed OpenAI in November, ranking in the top 3 in LiveBench’s global large language model performance rankings. China’s low-cost models, which were developed at a cost well below the cost of their competitors, took two places in the top 15.
– Robert Lea and Jasmine Lyu, analysts
Click here for research.
DeepSeek, known for its unconventional innovations, published its latest paper this week via the open data repository arXiv and the open source platform Hugging Face. The article lists 19 authors, with Liang’s name appearing last.
Continuously driving DeepSeek’s research agenda, the founder pushed his team to rethink how large-scale AI systems are designed and built.
The latest research addresses challenges such as training instability and limited scalability, noting that the new method involves “rigorous infrastructure optimization to ensure efficiency.” Based on ByteDance Ltd.’s 2024 survey of hyperlink architectures, tests were conducted on models ranging from 3 billion to 27 billion parameters.
The technique shows promise “for the evolution of basic models,” the authors said.
This article was generated from an automated news agency feed without modifications to the text.
