Is DeepSeek's New Model Coming? Series of Technical Papers Released in China

Second Paper Released This Year

The development team behind China's artificial intelligence (AI) model "DeepSeek" has been publishing a series of technical papers since the beginning of this year.

According to Chinese media outlets such as China Business Network on January 14, the DeepSeek research team, including founder Liang Wenfeng, released an online paper on January 12 titled "Scalable Retrieval-Based Conditional Memory: A New Axis of Sparsity for Large Language Models (LLMs)."

In this paper, co-authored with researchers from Peking University, the DeepSeek team presented a technical approach that separates "conditional memory" and "computation" in large models, thereby reducing errors and saving computing power, according to the media reports.

Just as humans recall existing memories when encountering familiar knowledge without new reasoning, the approach allows LLMs to avoid repetitive deep reasoning for frequently encountered information.

This paper was published just 11 days after DeepSeek released another paper on the "Manifold-Constrained Hyperconnection" framework on January 1. Given that DeepSeek has previously released research findings ahead of major model announcements, there is speculation that the unveiling of its next-generation AI model may be imminent.