DeepSeek has released a new "competitor to ChatGPT and Claude". A year ago, it crashed the market

Chinese AI startup DeepSeek released a preview version of its big language model V4 on Friday, April 24, giving users a chance to test its new features and capabilities. DeepSeek calls the updated version of its AI model the most powerful open source platform, challenging rivals OpenAI's GPT-5.4 and Anthropic's Claude Opus 4.5.

The V4 version is customized to run on chips from tech giant Huawei, highlighting China's growing capabilities in the AI industry, Reuters notes. The close collaboration with Huawei to create the V4 contrasts with DeepSeek's past experience of relying on Nvidia technology, although the startup did not disclose which processors were used to train the new product.

V4 is available in two variants - a professional "Pro" version and a lighter and cheaper "Flash" version. The model is optimized to work with popular agent-based AI tools (programs that can independently perform tasks in a digital environment), such as Claude Code from Anthropic and OpenClaw. Similar to DeepSeek's previous V3 version, the new AI model is distributed open source, allowing developers to freely download it, run it in-house, and customize it for specific tasks.

DeepSeek claims that the V4 performs admirably against Chinese competitors, particularly succeeding in the agent-based tasks of executing complex instructions, processing knowledge arrays and logical data inference.

Among the architectural innovations, the startup highlighted Hybrid Attention Architecture technology, which improves the AI's "memory" during long dialogs, as well as the expansion of the context window to 1 million tokens, which makes it possible to process entire code bases or voluminous documents with a single request, Bloomberg reports.

DeepSeek, founded in 2023, attracted attention in late 2024 with its free AI model V3. In January 2025, the Chinese AI startup released the R1 model, an open-source neural network that mimics the human thought process, Bloomberg points out. It not only equaled in tests with American market leaders, but also caused panic among investors with the details of its development, provoking a trillion-dollar sell-off in the U.S. stock market. According to statements by the developers of R1, its creation took only two months and less than $6 million. For comparison - training analogs from OpenAI or Google, according to some 2024 data, cost from $100 million to several billion dollars, which cast doubt on the expediency of colossal spending by Western corporations on AI infrastructure, recalls CNBC.

The new V4 model is unlikely to create the same furor as the R1 version, as the market is already used to the competitiveness and cheapness of Chinese neural networks, MorningStar senior analyst Ivan Su told CNBC in a commentary. In addition, DeepSeek is now directly competing with other Chinese open source AI models - from Alibaba and ByteDance, CNBC adds. "This kind of positioning [in the AI market] didn't exist in the days of R1, and that in itself speaks to how much internal competition has intensified," Ivan Su noted (quoted by CNBC).

The release of a new DeepSeek model caused shares of Chinese chipmakers to surge on April 24 as investors expect demand for Chinese AI hardware to rise, Bloomberg notes. SMIC shares rose 9.4 percent in Hong Kong and Hua Hong Semiconductor shares soared more than 13 percent, while model rivals such as Zhipu lost about 8 percent in value. Now DeepSeek, which is not a public company, is negotiating its first round of funding with Alibaba and Tencent, and its market valuation could exceed $20 billion, Bloomberg notes.

It is noteworthy that the release of V4 took place against the backdrop of another aggravation of U.S.-China relations in the field of AI: on April 23, the White House accused China of stealing intellectual property of U.S. laboratories on an "industrial scale", writes Reuters. DeepSeek found itself at the center of this scandal: Washington accuses the company of violating U.S. export controls when purchasing advanced Nvidia chips, while Anthropic and OpenAI claim improper use of their proprietary (closed-code language models) data for training, the agency specifies.

DeepSeek acknowledges the use of Nvidia hardware but has not commented on the issue of export restrictions, insisting that its models are trained on open-source data and do not use competitors' data. The Chinese Embassy in Washington called the allegations baseless, emphasizing Beijing's commitment to protecting intellectual property rights.

On April 24, Huawei, whose Ascend line of chips plays a key role in China's drive to reduce its reliance on U.S. semiconductor technology, confirmed a tight partnership with DeepSeek to power V4 across its entire line of high-performance systems. However, it remains unclear what proportion of Huawei's chips were used compared to solutions from Nvidia, CNBC adds. Due to US export restrictions imposed in 2022, Chinese developers cannot directly buy Nvidia's most advanced gas pedals. At the same time, Beijing is actively pushing companies to switch to domestic alternatives, CNBC says.

This article was AI-translated and verified by a human editor