DeepSeek's New Blueprint: Train Smarter AI for Less Money

Chinese AI company DeepSeek has started 2026 by publishing a new research paper. It proposes a major change to how the most powerful AI models are built. The paper introduces a method called Manifold-Constrained Hyper-Connections, or mHC. This technical approach rethinks the core architecture—the fundamental design—used to train "foundational" AI models. These are the large systems that power advanced chatbots and image generators. For DeepSeek, the goal is cost reduction. The Hangzhou-based startup aims to train bigger, more capable AI models while spending less on computing power. This strategy is critical for competing with better-funded U.S. rivals like OpenAI and Anthropic. The research is co-authored by DeepSeek's founder, Liang Wenfeng. It signals the company's continued focus on achieving more with fewer resources, a key challenge in the expensive global AI race.