DeepSeek 在 2026 年 1 月 2 日发布论文，提出名为“Manifold-Constrained Hyper-Connections（流形约束超连接）”的新框架，并通过 arXiv 与 Hugging Face 公开。作者称其目标是在提升可扩展性的同时，降低训练先进 AI 系统所需的算力与能耗，体现在缺乏对英伟达芯片自由获取的情况下，中国 AI 产业对效率的持续追求。

这类论文过去曾预示 DeepSeek 的重大模型发布：这家杭州初创公司在约一年前推出 R1 推理模型，成本仅为硅谷竞争对手的一小部分。公司随后发布了多款较小平台，但市场对下一代旗舰系统“R2”的期待升温，预计将在 2026 年 2 月春节前后出现。美国对先进半导体的限制迫使研究者探索更非常规的方法与架构。

彭博行业研究认为，R2 可能在未来几个月发布，并有潜力再次搅动全球 AI 格局。LiveBench 排名显示，谷歌 Gemini 3 在 11 月超越 OpenAI，进入全球大模型性能前三；中国的低成本模型占据了前十五名中的两个席位。DeepSeek 最新论文列出 19 位作者，创始人梁文锋署名在末位；实验覆盖 30 亿到 270 亿参数规模，并在字节跳动 2024 年关于超连接架构研究的基础上推进，以改善训练不稳定与可扩展性受限等问题。

On Jan. 2, 2026, DeepSeek published a paper, released via arXiv and Hugging Face, describing a new framework called Manifold-Constrained Hyper-Connections. The authors say it aims to improve scalability while cutting the computation and energy required to train advanced AI systems, highlighting China’s push for efficiency as it competes with OpenAI despite limited access to Nvidia chips.

DeepSeek’s technical publications have previously foreshadowed major releases: about a year ago, the Hangzhou startup surprised the industry with its R1 reasoning model, built at a fraction of Silicon Valley rivals’ cost. It has since shipped smaller platforms, but attention is rising for the next flagship, widely dubbed R2, expected around the Spring Festival in February 2026. US curbs on leading semiconductors are pushing researchers toward unconventional methods and architectures.

Bloomberg Intelligence says R2 could arrive in the next few months and may disrupt the global AI sector again. LiveBench rankings show Google’s Gemini 3 overtook OpenAI in November to reach a top-three slot, while China’s low-cost models held two of the top-15 positions. DeepSeek’s paper lists 19 authors with founder Liang Wenfeng last; tests span 3B to 27B parameters and build on ByteDance’s 2024 work on hyper-connection architectures to address instability and limited scalability.