Chinese artificial intelligence startup DeepSeek on Friday released preview versions of its latest AI model, stepping up competition with leading US tech firms.
The new V4 model has been highly anticipated, with users keen to compare it with rivals such as ChatGPT by OpenAI, Claude from Anthropic and Gemini by Google.
The company introduced “pro” and “flash” versions of the open-source V4 model, saying they offer major improvements in knowledge, reasoning ability and “agentic” functions, which allow the system to complete complex tasks on its own.
DeepSeek also said the new model partly runs on chips developed by Huawei, reducing dependence on US chipmaker Nvidia.
V4 follows the earlier V3 model released in late 2024. However, it was DeepSeek’s reasoning-focused R1 model, launched in January 2025, that drew global attention for being more cost-effective than similar systems from US firms.
The company claims its top version, “V4 Pro Max,” performs better than OpenAI’s GPT-5.2 and Google’s Gemini 3.0-Pro in several reasoning tests, though it falls slightly behind their more advanced versions. The launch came shortly after OpenAI introduced its GPT-5.5 model.
DeepSeek also said its V4 “pro” version could outperform Claude’s Sonnet 4.5 and come close to the performance of Claude’s Opus 4.5 model in handling complex tasks. The “flash” version is designed to deliver similar results in simpler tasks with faster performance.
Analysts say the new model could be a strong competitor.
“Based on early results, DeepSeek V4 looks highly competitive with US rivals,” said Lian Jye Su of research firm Omdia.
Marina Zhang of the University of Technology Sydney described the launch as an important step for China’s AI sector as global competition grows.
DeepSeek offers free chatbot services on web and mobile platforms and promotes its technology as open source, allowing developers to modify and build on it.
The company said both versions of V4 can process up to 1 million tokens of information at a time, a significant increase from the 128,000-token limit in V3.
Huawei confirmed its Ascend chips are compatible with the new model, highlighting China’s efforts to reduce reliance on US technology.
A January report from Microsoft showed DeepSeek is gaining popularity in developing countries, especially where Huawei devices are widely used.
However, some experts remain cautious. Ivan Su of Morningstar said the new model is a solid update but not as groundbreaking as the earlier R1 release.
Competition in China’s AI market has also increased, while independent testing is still needed to verify DeepSeek’s performance claims.
Meanwhile, tensions remain over accusations from US companies. Anthropic and OpenAI have alleged that DeepSeek used their models’ outputs to improve its own systems, a claim also echoed by US officials.
China has rejected the allegations, calling them unfair attempts to limit the growth of its technology companies.