China's DeepSeek kicked off 2026 with a new AI training method that analysts say is a 'breakthrough' for scaling

DeepSeek, a prominent Chinese AI startup, has kicked off the year with an innovative approach to training artificial intelligence that analysts are calling a 'significant breakthrough.' The company released a research paper on Wednesday detailing their new method for training large language models, which they believe could redefine the trajectory of foundational AI models. The paper, co-authored by founder Liang Wenfeng, introduces a novel technique known as "Manifold-Constrained Hyper-Connections" (mHC). This approach allows for enhanced internal communication within models while maintaining stability and computational efficiency as models scale. In the context of growing language models, researchers traditionally attempt to boost performance by enabling different components of a model to share information more freely. However, this often leads to instability. DeepSeek's latest findings suggest that by constraining how models share this information, they can achieve richer interactions without compromising stability. Wei Sun, principal analyst for AI at Counterpoint Research, praised the new methodology, describing it as a 'striking breakthrough.' She emphasized that DeepSeek's approach cleverly combines various techniques to reduce the costs typically associated with training models. Despite a slight increase in costs, the potential for enhanced performance is substantial. Sun noted that the publication reflects DeepSeek's internal capabilities and their commitment to merging rapid experimentation with innovative research ideas. This advancement could allow DeepSeek to overcome computational bottlenecks and achieve significant leaps in AI intelligence, reminiscent of their 'Sputnik moment' in January 2025 when they launched their R1 reasoning model, which notably challenged established competitors. Lian Jye Su, chief analyst at technology research firm Omdia, stated that DeepSeek's research could influence the broader industry, inspiring rival AI labs to explore similar methodologies. He remarked on the company's willingness to share important findings, indicating a newfound confidence within the Chinese AI sector. As DeepSeek gears up for the anticipated launch of its next flagship model, R2, the timing of this paper is noteworthy. Initially expected in mid-2025, the launch has faced delays due to performance issues and shortages of advanced AI chips, which have complicated the development and deployment of cutting-edge models in China. While the paper does not specifically mention R2, its release has raised questions, especially given DeepSeek's history of publishing foundational research ahead of significant model launches. Analysts believe that the novel architecture introduced could play a critical role in DeepSeek's future developments, although some remain cautious about the timeline and the potential for a standalone R2 model given recent updates to their existing R1 model.