DeepSeek unveils new AI reasoning method amid anticipation for its next-gen model

Chinese artificial intelligence (AI) startup, DeepSeek, has unveiled a new reasoning method as the market continues to anticipate the release of its next-gen model. According to reports, the company has just introduced a novel approach that will help the reasoning capabilities of large language models (LLMs), as the public awaits the release of its new model.

According to reports, a paper published recently revealed that the feat was achieved via a collaboration between researchers at Tsinghua University and the AI startup. According to the paper, DeepSeek was able to develop a technique that combines methods referred to as generative reward modeling (GRM) and self-principled tuning. The dual approach will enable LLMs to deliver better and faster results to general queries.

DeepSeek reveals new AI reasoning method

According to the researchers, the resulting DeepSeek-GRM models were able to outperform existing methods, achieving a competitive performance with strong public reward models. Reward modeling is a process that helps an LLM to determine human preferences. The researchers mentioned that the AI startup intends to make the GRM models open source, but the company has yet to give a timeline for the development.

The published academic research paper is coming after speculations about the potential next move of the startup after enjoying global attention it gathered from its DeepSeek V3 foundation model and its R1 reasoning model. At the time, the R1 model, which was released earlier this year, gained prominence after it was discovered to have performed better than some older models, including OpenAI’s first ChatGPT model.

The release of the R1 model also caught Silicon Valley off guard, considering the fact that the model was open source and was completed with a fraction of the costs used in other AI models. While the attention seems to be slowing down, there has been some remarkable progress among rival firms in the AI industry, showing that the rivalry in the sector is one for the ages.

However, a Reuters report last month mentioned that DeepSeek-2, the successor to the globally celebrated R1 might be in the works. According to the report, the company could be looking to release it towards the end of this month, as it looks to capitalize on its growing profile in the AI industry. DeepSeek, on the other hand, has remained tight-lipped about the R2 model, maintaining silence on official public channels, with a customer service account denying the rumor with business clients, according to reports.

Development on the V3 model continues

DeepSeek, based in Hangzhou, was founded in 2023 by entrepreneur Liang Wenfeng and has been in the global spotlight for the right reasons in the last few months. However, the company has refused to make public communications, deciding to focus all its time and energy on research and development. While it remains to be seen what the company has for the AI industry, it has surely been cooking, thanks to rumors.

Last month, the company announced a development it has been working on, showing its upgraded V3 model named DeepSeek-V3-0324 to the world. According to the firm, the upgraded model offered “enhanced reasoning capabilities, optimized front-end web development, and upgraded Chinese writing proficiency.” In February, the company out-sourced five of its code repositories, allowing developers to review and contribute to its software development. The startup mentioned that it would make “sincere progress with full transparency.”

In the same month, Liang published a technical study on native sparse attention, a method used to improve the efficiency of LLMs in processing large amounts of data. The 40-year-old Liang is also the founder of DeepSeek’s parent firm, High-Flyer Quant, the hedge fund that has provided the financial prowess to fund the AI startup’s technical advances. In late February, the entrepreneur was part of a symposium that assembled tech entrepreneurs, with Chinese President Xi Jinping hosting them. The President hailed the startup as a sign of the country’s resilience in the face of the United States’ moves to contain China’s AI progress.

Cryptopolitan Academy: Coming Soon - A New Way to Earn Passive Income with DeFi in 2025. Learn More