DeepSeek reportedly will not share its upcoming AI models with US engineers as of February 26, 2026, instead granting early access to Chinese companies, further escalating the tech war between the US and China.
Null Photo | Null Photo | Getty Images
Chinese artificial intelligence startup DeepSeek on Friday released a preview version of its long-awaited V4 large-scale language model, allowing users to test its new features.
This release comes more than a year after DeepSeek introduced its R1 inference model, which shook up the global technology market with its incredible performance and cost efficiency.
Like DeepSeek’s previous V3 model, the latest upgrade is open source, allowing developers to download the code, run it locally, and modify it.
The Hangzhou-based company claimed that V4 has achieved superior performance against domestic competitors, especially in agent-based tasks, knowledge processing, and reasoning.
The company added that DeepSeek-V4 is optimized for use with popular agent tools such as Anthropic’s Claude Code and OpenClaw.
This model is available in both “Pro” and “Flash” versions depending on size.
DeepSeek was founded in 2023 and gained attention in late 2024 with its free and open source V3 model. This model is trained using a less powerful chip and uses OpenAI and google.
A few weeks later, in January 2025, the company released its inference model R1, which achieved similar benchmarks or outperformed many of the world’s leading LLMs.
The R1 model alarmed investors when DeepSeek revealed that it took just two months and less than $6 million to build the model using low-capacity Nvidia chips. This called into question America’s lead in AI and Big Tech’s massive spending on AI infrastructure.
Since then, DeepSeek has released a series of model upgrades, but none have matched the impact of R1.
Morningstar senior equity analyst Ivan Su told CNBC that V4 is unlikely to have the same impact as R1 because the market has already factored in the reality that Chinese AI will be competitive and low cost to use.
However, with DeepSeek’s latest positioning, other Chinese open source models are direct competitors, Su said.
“This is a framework that did not exist in R1, and this alone shows how intense domestic competition has become,” he added.
Since the release of R1, DeepSeek has faced increased competition in China’s fast-growing AI space, with players including: alibaba ByteDance will also release new models this year.
Shares of several other Chinese AI companies fell in Hong Kong trading on Friday. Knowledge Atlas Technology, also known as MiniMax and Zhipu, each fell about 8%, while Hangzhou-based developer Manycore Tech fell 9%.
Which chip trained V4?
A big question after the release of DeepSeek’s V4 model is which chip was used to train the model.
Chinese technology giant Huawei on Friday confirmed that its latest AI computing cluster powered by Ascend AI processors can support DeepSeek’s V4 model.
However, it is unclear to what extent Huawei’s Ascend chips were used to train the model compared to Nvidia’s chips.
U.S. export controls restrict Chinese AI developers from directly purchasing Nvidia’s cutting-edge chips for AI training.
Meanwhile, the Chinese government has stepped up efforts to develop the domestic chip industry, reportedly encouraging Chinese tech companies to adopt domestic alternatives from chipmakers such as Huawei over foreign alternatives.
Shares of the Chinese contract chipmaker rose in Hong Kong after DeepSeek announced its V4 release. SMIC and Huahong Semiconductor This rapidly increased to 8.9% and 15.2%, respectively.
