Report: DeepSeek prefers new AI model and wants to release R2 before May

Contents

The Chinese AI company DeepSeek apparently wants to benefit from the current hype surrounding its AI model R1 and release its successor R2 earlier than planned. It is now being reported that DeepSeek originally planned to release R2 at the beginning of May, but wants to release the new AI model as early as possible. A flat company structure is likely to help, as DeepSeek is organized differently to Chinese companies. R2 should provide better support for programming and be able to discuss in more languages than just English.

DeepSeek published R1 as a new AI language model with reasoning back in November 2024. Like OpenAI's o1, DeepSeek R1 takes more time to answer, but is said to be more accurate than other AI models in many cases. It is also open source and was apparently trained with less effort. This has led to the formation of crisis teams at competitors and large tech companies. A look behind the scenes of DeepSeek's R1 reasoning model shows how the model works and what it means for AI development.

DeepSeek founder stays in the background

However, DeepSeek's AI models are not only impressing the competition, but also the US government. After all, it has declared leadership in artificial intelligence (AI) to be a national priority. DeepSeek's continued success, even with the new AI model R2, is likely to further mobilize Chinese authorities and companies. Dozens have already started to integrate DeepSeek AI into their products and services, writes Reuters.

So far, however, little is known about DeepSeek and its founder Liang Wenfeng, who rarely makes public appearances. Liang is described by former employees as “reserved and introverted” and last gave an interview in July 2024. According to in-depth research by Reuters, DeepSeek operates more like a research laboratory than a profit-oriented company.

DeepSeek structure atypical for China

The AI company dispenses with the traditional hierarchies of Chinese companies and relies on a flat management structure. DeepSeek also avoids the typical pressure of Chinese tech companies on low-paid employees and the “996” rule: working from 9am to 9pm 6 days a week. Former employees speak of 8-hour days in a collaborative atmosphere. Employees are also said to be paid almost twice as much as their competitors.

Even before DeepSeek caused a worldwide stir, there were signs that the Chinese government had developed an interest in the AI company. Back in January, DeepSeek founder Liang represented the AI sector at a meeting with China's Prime Minister Li Qiang in Beijing. It is remarkable that he was chosen ahead of much better-known and larger companies.

DeepSeek's success delights Beijing

DeepSeek's subsequent success due to its cost-effective AI model has led the Chinese government to believe that it can now more than keep up with the USA in terms of innovation. At least 13 city governments and ten state-owned energy companies have so far announced that they have integrated DeepSeek into their systems. In addition, well-known companies such as Lenovo, Baidu, and Tencent have integrated or are currently integrating these AI models into their products.

However, there is criticism of DeepSeek AI in Western countries. The Chinese open-source AI Deepseek V3 is censored and hides many questions about China's politics. There are also data protection concerns: following Italy, South Korea has also banned DeepSeek to protect personal data. In Germany, too, authorities and IT security experts have expressed serious security concerns about DeepSeek. The apparently very extensive storage of user data is also being criticized in this country.

(fds)