?? DESCRIPTION

Chinese AI startup DeepSeek on Friday unveiled a preview of its long-awaited V4 model while also moving to raise outside funding for the first time, developments that rattled some Chinese AI stocks, lifted chipmaking shares across Hong Kong and mainland markets, and renewed questions about which chips powered the new release.

The Hangzhou-based company released the V4 as a test version, giving developers early access to try out its features. Like its predecessor, the V3 model, V4 is open source, meaning developers can download, run, and change the code on their own systems. The model comes in two sizes, a “pro” version and a smaller “flash” version.

DeepSeek said V4 performs well against domestic rivals, particularly in tasks involving AI agents, knowledge handling, and inference. The company also said the model has been built to work with popular agent tools, including Anthropic’s Claude Code.

The release arrives more than a year after DeepSeek’s R1 reasoning model shook global tech markets. When R1 came out in January 2025, it matched or beat many leading AI models, and DeepSeek revealed it had taken just two months and less than $6 million to build, using lower-grade Nvidia chips. That disclosure rattled investors and raised questions about the U.S. lead in AI as well as Big Tech’s massive spending on AI infrastructure.

The company now faces growing competition in China’s booming AI sector. Alibaba and ByteDance are among the players that have released new models this year. On Friday, the V4 release sent shares of several Chinese AI companies lower in Hong Kong.

Zhipu AI fell around 8-9%, MiniMax dropped roughly 7-8%, and Manycore Tech slid 9%.

Chipmaking stocks, however, moved in the opposite direction as the V4 release drove optimism over AI-driven demand.

Semiconductor Manufacturing International Corp, the country’s largest chipmaker by volume, jumped 11% in Hong Kong, while Hua Hong Semiconductor rallied more than 18%.

On the mainland, Cambricon Technologies and Moore Threads Technology each gained between 4% and 6%, and Hygon Information Technology climbed more than 10%.

Which chips trained DeepSeek V4?

One of the biggest questions following the release is what hardware DeepSeek used. According to Reuters, Huawei confirmed Friday that its Ascend 950-based supernode can support the V4 model and said its full line of high-performance systems now works with the V4 series.

However, DeepSeek did not say which chips it used to train the model, leaving the question unanswered.

Chinese AI developers have been blocked from buying Nvidia’s most advanced chips because of U.S. export controls that began in 2022. Beijing has since pushed its tech companies toward domestic alternatives from chipmakers such as Huawei.

The V4 launch came one day after the White House accused China of stealing U.S. AI labs’ intellectual property on an industrial scale, a charge that could strain relations ahead of a planned summit between U.S. and Chinese leaders next month.

DeepSeek has been at the center of that dispute, with Washington alleging it obtained restricted Nvidia chips and with companies including Anthropic and OpenAI saying it improperly copied their proprietary models. The Chinese Embassy in Washington rejected what it called “baseless allegations.”

Fundraising to hold onto researchers

As reported by Cryptopolitan previously, DeepSeek is in talks with a small group of strategic investors, including Tencent and Alibaba, about raising funds at a valuation above $20 billion, its first outside fundraising.

The expected amount is in the low hundreds of millions of dollars, far below the billions typically raised by peers. Moonshot, which runs the Kimi AI models, was last valued at $18 billion, while MiniMax and Zhipu carry valuations of $34 billion and $58 billion, respectively.

The fundraising is not being driven by an urgent need for cash but mainly to retain researchers, sources told FT. Some of the researchers have left for rivals whose valuations have soared over the past year. Stock options make up a large part, if not the majority, of an AI researcher’s salary, and without a clear valuation, DeepSeek has struggled to compete.

Guo Daya, a lead author of the R1 paper, joined ByteDance, while Wang Bingxuan, a veteran of DeepSeek’s model training team, left for Tencent.

Founder Liang Wenfeng, who has funded the company through his quantitative trading firm, is also considering other options to establish a valuation, including a share buyback or a performance-based valuation method, in case fundraising terms cannot be reached.

The smartest crypto minds already read our newsletter. Want in? Join them.

This detailed match analysis covers key moments, player performances, and tactical insights. Download the full report in PDF format.