Chinese AI startup DeepSeek has recently developed an advanced reasoning method aimed at boosting the performance of large language models (LLMs). This innovative approach, created in collaboration with researchers from Tsinghua University, merges generative reward modeling (GRM) with self-principled tuning, significantly enhancing both the speed and quality of outputs generated by LLMs.
A paper published by the company highlights that the DeepSeek-GRM models achieved impressive results, at times surpassing existing public reward models. Although DeepSeek plans to open-source these GRM models, no specific release date has been communicated yet.
This announcement occurs amid rising anticipation regarding DeepSeek’s next significant product launch. Earlier this year, the startup gained international acclaim with its R1 reasoning model, which surpassed several older models, including the original ChatGPT.
The R1 model was not only recognized for its superior performance but also for being open-source and developed on a relatively low budget. Reports suggest that the company is gearing up to introduce its R2 model imminently, potentially by the end of April, despite no official comments from the company confirming the timeline.
Founded in 2023 in Hangzhou by entrepreneur Liang Wenfeng, DeepSeek has emphasized research over public visibility, steadily gaining traction within the AI community. The startup recently presented the DeepSeek-V3-0324, featuring enhanced reasoning, superior web development capabilities, and improved Chinese writing functionality. Additionally, DeepSeek has made parts of its codebase public, reflecting its commitment to open development practices.
Supported by High-Flyer Quant, a hedge fund led by Liang, the startup is positioning itself as a formidable player in the global AI landscape, receiving commendations from China’s President Xi Jinping for its innovative contributions and strategic importance.
