China's DeepSeek kicked off 2026 with a new AI training method that analysts say is a 'breakthrough' for scaling (www.businessinsider.com)

China's DeepSeek has unveiled a new AI training method called "Manifold-Constrained Hyper-Connections" (mHC), which aims to enhance the scalability of large language models while maintaining stability and computational efficiency. Analysts, including Wei Sun from Counterpoint Research, describe this approach as a "striking breakthrough," suggesting it could significantly influence the development of foundational models in the AI/ML landscape. By enabling more comprehensive internal communication among various components of a model without instability, DeepSeek is positioning itself to surpass compute bottlenecks and deliver higher performance at lower costs. The timing of this announcement is noteworthy as DeepSeek is reportedly preparing to launch R2, its next flagship model, following delays due to performance concerns and chip shortages. The insights from the published research may not only refine DeepSeek's upcoming models but could also inspire rival AI labs to adopt similar techniques, signaling a collaborative spirit within the industry. Analysts believe that DeepSeek's willingness to share innovative findings showcases a newfound confidence in the Chinese AI sector, potentially reshaping competitive dynamics against leading players like OpenAI and Google, especially in the Western market.