Main points
- DeepSeek has released a new language model V4 with two versions, Pro and Flash, which differ in the number of parameters and a context window of up to 1 million tokens.
- The V4 model has raised concerns over accusations of plagiarism and privacy issues, as the data is stored on Chinese servers accessible to Chinese government agencies.
DeepSeek V4 is a game changer / Depositphotos
DeepSeek has released a preview of its new large language model V4, which promises significant increases in power and efficiency. The update includes two versions and demonstrates the ambitions of Chinese developers in the global race for leadership in artificial intelligence.
What does DeepSeek V4 offer?
Chinese startup DeepSeek, which has already caused a real sensation in global markets, has officially announced the launch of the Preview version of its new flagship model DeepSeek V4. This release was a response to the intensifying competition between China and the United States in the field of high technology. The new series is presented in two versions: Pro and Flash, both of which are already available to users on the company's official website, writes Tech Xplore.
The technical specifications of the new model, as listed in the DeepSeek announcement, are impressive in their scale:
- The Pro version has a total number of parameters reaching 1.6 trillion, of which 49 billion are active.
- In turn, the lighter version of Flash works on the basis of 284 billion parameters, of which 13 billion are active.
- One of the most important achievements of the developers was the expansion of the context window to 1 million tokens for both variants. This allows the neural network to process and remember huge amounts of information at once, which is a significant improvement over the previous V3 model, which supported only 128,000 tokens.
DeepSeek representatives claim that the Pro version demonstrates exceptional abilities in logical thinking, mathematics and programming, outperforming all open-source competitor models, writes Neowin. According to them, in logic tests, the V4 Pro Max shows better results than OpenAI GPT-5.2 and Gemini 3.0-Pro from Google, although it is slightly inferior to GPT-5.4 and Gemini 3.1-Pro.
Special attention is paid to “agent” capabilities – the ability of the system to autonomously perform complex workflows. In this aspect, the Chinese development can surpass Claude Sonnet 4.5 and approach the level of Claude Opus 4.5.
DeepSeek V4 Results / Photo DeepSeek Results
The Flash version is presented as the most profitable solution for developers. It demonstrates logical capabilities that are almost as good as the Pro version, and copes with simple standalone tasks on an equal footing.
- The cost of using the Flash API is very low: input costs between $0.028 and $0.14, and output costs $0.28.
- For the Pro model, prices are higher: incoming data will cost $0.145 – $1.74, and outgoing data will cost $3.48.
Slow down with usage: what's the problem with DeepSeek
Steal stolen goods
Despite its technical successes, the release of DeepSeek V4 has been accompanied by geopolitical tensions and accusations of plagiarism. American companies Anthropic and OpenAI accuse the Chinese startup of illegally using their technologies. In particular, it is about the distillation method, when a less powerful model is trained on the results of the work of a stronger system.
US administration officials have also expressed concern that foreign companies are exploiting American innovations, a claim China has dismissed as unfounded pressure on business.
Privacy issues
The use of DeepSeek also comes with serious security and privacy concerns, which have been the subject of international investigations since its initial release. The main problem is that the developer, Hangzhou DeepSeek Artificial Intelligence, is based in China and explicitly states that the data is stored on Chinese servers, making it accessible to Chinese government agencies.
Under Chinese law (including national security and cybersecurity laws), companies are required to provide user data to law enforcement and government agencies upon request. Cybersecurity researchers have discovered code that allows data to be transmitted directly through the infrastructure of state-owned companies like China Mobile.
National security experts in the US and other countries classify DeepSeek as a potential tool for cyberespionage and influence operations.
Technical gaps
- In January 2025, researchers from Wiz Research discovered the leak of over a million records (chat history, API keys, internal logs) through a misconfigured cloud database that was publicly accessible.
- NowSecure experts recorded that the DeepSeek app for iOS was forcibly disabling the App Transport Security (ATS) security system, which is designed to encrypt data during transmission.
- Cisco tests have shown that the R1 model practically does not block malicious requests (jailbreak attacks), which allows it to be used to write malicious code or generate dangerous content.
So if you plan to use DeepSeek, do not provide personal data, corporate secrets, or confidential documents. The safest option for businesses is to self-host (locally deploy) the open model on your own servers so that the data does not leave your infrastructure.