Hong Kong – DeepSeekthe Chinese artificial intelligence startup that shook global markets last year, released previews of its latest big update on Friday, as the rivalry of the Artificial intelligence (AI) between China and USA it gets hot.
DeepSeek V4 has been highly anticipated by users, who wanted to compare its performance with that of American competitors such as ChatGPT of OpenAIClaude de Anthropic and Gemini of Google. Anthropic and OpenAI have accused DeepSeek of unfairly basing its technology on theirs.
Some industry analysts expected the new model to arrive more than two months early, at the beginning of the Lunar New Year.
DeepSeek claims that the new V4 open source models, which include “pro” and “flash” versions, present great improvements in cognition, reasoning and their “agent” capabilities, that is, the ability to perform complex tasks and workflows autonomously. Another big change is that they rely in part on computer chips made by Chinese tech giant Huawei, reducing DeepSeek’s dependence on American chipmakers like Nvidia.
V4 is the successor to V3, an AI model that DeepSeek launched in late 2024.
But it was DeepSeek’s specialized “reasoning” AI model, called R1, that took markets by surprise with its launch in January 2025. DeepSeek claimed it was more profitable than OpenAI’s similar model and became a symbol of how China was catching up with the United States in technological advances.
DeepSeek claims that the “V4 Pro Max” version is “outperforming” in terms of standard reasoning tests relative to OpenAI’s GPT-5.2 model and Google’s Gemini 3.0-Pro. It falls “slightly” below GPT-5.4 and Gemini 3.1-Pro, according to the company. DeepSeek’s publication came hours after OpenAI unveiled its new GPT-5.5 model on Thursday.
In terms of “agent” capabilities, the Chinese company claims that the “pro” V4 version could surpass Claude’s Sonnet 4.5 and is close to the level of Claude’s Opus 4.5 model according to its own assessment.
The “flash” version of V4 performs on par with the “pro” version in simple agent tasks and has reasoning capabilities that are very close, according to DeepSeek.
“Based on the benchmark results, it looks like DeepSeek V4 is going to be very competitive against its US rivals,” says Lian Jye Su, chief analyst at technology research and advisory group Omdia.
Marina Zhang, associate professor at the University of Technology Sydney, said the launch of DeepSeek V4 is a “critical milestone for the Chinese AI industry”, especially as global competition intensifies in the quest for self-sufficiency in critical technologies.
DeepSeek offers a free chatbot for web and mobile. Unlike the best models from Anthropic, Google and OpenAI, it describes its technology as “open source” in the sense that it allows developers to access it to modify and build on its core technology.
Both the “pro” and “flash” versions of V4 have a context window of 1 million tokens, a parameter of the amount of information an AI model can process and remember, and operate more efficiently, according to the startup. This is a significant improvement over the previous version, as V3 supported a context window of 128,000 tokens.
Huawei said in a separate statement Friday that its Ascend chips and related technology are compatible with DeepSeek V4 models. It’s a demonstration of the technical feasibility of operating outside the Nvidia-dominated computing ecosystem “amid sustained technological decoupling between China and the United States,” Zhang said.
A Microsoft report from January showed that the use of DeepSeek has been gaining ground in many developing countries, especially those where the use of Huawei phones is widespread.
However, some analysts are skeptical. Ivan Su, senior equity analyst at Morningstar, says that while V4 is a “competent” follow-up, it’s not as big a step forward as the launch of R1.
“Domestic competition has intensified significantly since the launch of R1,” Su said. “Versus US models, DeepSeek’s own assessment suggests its capabilities are broadly consistent on most fronts, but independent assessments are needed before drawing definitive conclusions.”
In February, Anthropic accused DeepSeek and two other China-based AI labs of “industrial-scale campaigns” to “illicitly mine Claude’s capabilities to improve their own models.” It said that they did it through a technique called distillation, which “consists of training a less capable model with the results of a stronger one.” OpenAI made similar accusations in a letter to US lawmakers.
This week, Michael Kratsios, US President Donald Trump’s top science and technology adviser, also accused foreign tech companies “primarily based in China” of distilling major US AI systems and “exploiting American expertise and innovation.”
The Chinese embassy in Washington responded to the accusations by calling them “unjustified suppression of Chinese companies by the United States.”