Chinese AI startup DeepSeek has introduced a new reasoning method aimed at enhancing the performance of large language models (LLMs).

The approach, developed in partnership with researchers from Tsinghua University, combines generative reward modelling (GRM) and self-principled tuning to improve the speed and quality of LLM outputs.

According to a recently published paper, the resulting DeepSeek-GRM models achieved competitive results, even outperforming public reward models in some instances. Although DeepSeek has expressed plans to open-source the GRM models, no release date has been confirmed.

The announcement comes amid growing speculation about DeepSeek's next major release. The company gained global recognition earlier this year with its R1 reasoning model, which outperformed some older models like the original ChatGPT.

The R1's success was notable not just for its performance but also for being open source and developed on a relatively modest budget. Industry observers believe DeepSeek is preparing to unveil the R2 model soon, possibly by the end of the month, though the company has declined to comment officially.

Founded in Hangzhou in 2023 by entrepreneur Liang Wenfeng, DeepSeek has prioritised research over public relations, quietly building momentum in the AI sector.

The company recently showcased DeepSeek-V3-0324, an upgraded model with improved reasoning, better web development capabilities and enhanced Chinese writing. DeepSeek has also made parts of its codebase available to the public, signalling a commitment to open development.

Backed by High-Flyer Quant, Liang's hedge fund, the startup is emerging as a serious contender in the global AI race, drawing praise from the president of China, Xi Jinping, for its innovation and strategic significance.