DeepSeek sparks curiosity.

DeepSeek is focusing on research and development of new models, rather than appearing frequently in the media. Photo: SCMP .

In collaboration with researchers from Tsinghua University, DeepSeek has introduced a new method to improve the inference capabilities of large language models (LLMs). The method, published in a research paper on the evening of April 4th, helps LLMs produce better and faster results for common queries.

This technique combines two previously successful methods from DeepSeek. One is generative reward modeling (GRM), which allows the AI model to self-evaluate and refine its responses based on previous results, and the other is self-principled critique tuning.

Both methods rely on the "self-learning" aspect of AI, reducing reliance on direct human feedback or guidance, but with the aim of delivering results that are closer to human expectations.

According to researchers, despite being a new method, DeepSeek-GRM achieves outstanding results and competes with the most well-known and effective AI models currently available. DeepSeek plans to open-source GRM models, but no specific timeframe has been given.

After making a global impact with its V3 platform model and R1 inference model, DeepSeek published this academic paper on the online scientific archive arXiv, sparking curiosity about the company's next move.

Reuters predicts that DeepSeek-R2, the successor to R1, could launch in April, given the continued popularity of its predecessor. DeepSeek-R1 previously caused a global sensation in the tech world thanks to its superior performance relative to cost, making it competitive with current leading models.

DeepSeek has remained silent regarding the rumors. However, according to local sources, a DeepSeek customer service account denied the information in a group chat with enterprise clients.

Founded in Hangzhou in 2023 by entrepreneur Liang Wenfeng, DeepSeek has quickly garnered global attention in the past few months. But instead of capitalizing on its public fame, the company is focusing its resources on research and development.

Previously, DeepSeek upgraded its V3 model, releasing version DeepSeek-V3-0324. According to the announcement, this update features enhanced reasoning capabilities, optimization for front-end web user interface development, and improved Chinese writing skills.

In February, the startup also open-sourced five code repositories, affirming its commitment to "progress with full transparency." Also that month, the company announced a technical study on "native sparse attention," which helps improve the performance of LLMs in handling massive amounts of data.

DeepSeek is seen as a symbol of the dynamism of China's AI industry, at a time when the US is trying to curb the country's technological development.