DeepSeek on Friday published a preview of V4, a new version of its artificial intelligence model that the company says has been developed in close cooperation with Huawei and adapted to run on the Chinese firm’s Ascend processors. The Pro edition of V4, DeepSeek reported, ranks ahead of other open-source models on world-knowledge benchmarks and trails only Google’s closed-source Gemini-Pro-3.1.
The collaboration with Huawei marks a notable pivot for DeepSeek, which previously relied heavily on Nvidia GPUs during development. Huawei confirmed that its chips were used in portions of V4’s training workflow. Industry analysts interpreted the move as a concrete step toward greater domestic independence for China’s AI stack.
"This is a big deal for China’s AI industry," said He Hui, director of semiconductor research at consultancy Omdia. "Huawei’s Ascend chips are the country’s best homegrown alternative to Nvidia, and supporting DeepSeek V4 shows that top Chinese AI models can now run on Chinese hardware."
Most leading AI models are developed and deployed on Nvidia-made processors, and DeepSeek’s shift underlines longstanding concerns voiced by Nvidia that U.S. firms may lose parts of their developer ecosystem in China because of U.S. export restrictions and Beijing’s push for self-sufficiency. Nvidia CEO Jensen Huang was quoted this month saying on a podcast: "The day that DeepSeek comes out on Huawei first, that is a horrible outcome for our nation."
Developers and platform observers have taken note of V4’s rapid ascent on Hugging Face, a widely used repository and forum for machine learning models. Lewis Tunstall, a machine learning engineer at Hugging Face, said V4 is the fastest model to reach the top of the site’s trending list. He described the model as adept at processing extremely long and complex text tasks while operating at a substantially lower cost than leading alternatives, though he noted it currently lacks multimodal support for images and video.
DeepSeek’s emergence and performance have been controversial outside China. Several Western and some Asian governments have barred their institutions and officials from using DeepSeek, citing data privacy concerns. Washington and some U.S. competitors have also criticized DeepSeek’s rise as dependent in part on the misuse of American-derived know-how. DeepSeek has acknowledged prior use of Nvidia chips but has said it did not intentionally use synthetic data generated by OpenAI; the company has not clarified whether the Nvidia chips it used were subject to export bans.
The timing of Friday’s preview was notable. It followed a White House allegation made the previous day that China had engaged in large-scale intellectual property theft from U.S. AI laboratories, and it precedes a planned visit next month by U.S. President Donald Trump to meet China’s leader, Xi Jinping.
Trade and export issues remain unresolved. The U.S. government earlier this year permitted the sale of Nvidia’s H200 chips to China, but sources have said shipments have been delayed by disagreements over the terms of those sales in both countries.
Market responses to the V4 preview included a rally in domestic chipmakers on the expectation of broader adoption of homegrown processors. Huahong Semiconductor and SMIC both posted notable gains, rising 15% and 10% respectively. Nvidia shares also rose, buoyed in part by an optimistic revenue and profit forecast from Intel that reinforced investor confidence in continued AI-driven demand for compute.
Competition within China’s AI sector is intensifying. DeepSeek rocketed to national prominence a year ago, but the lead it enjoyed has narrowed as a wide array of domestic rival models has appeared. The release of V4 triggered sharp moves in rival stocks, with Zhipu AI and MiniMax each falling about 9% following the announcement.
DeepSeek positioned V4 as particularly suitable for AI agent applications, which require the capacity to orchestrate complex, multi-step operations and typically demand more compute than simple chatbot interactions. The company offers V4 in a Pro configuration and a lower-cost Flash edition; the preview phase is intended to gather developer feedback and refine the model before a final release. DeepSeek did not provide a timeline for finalizing V4.
Independent testing and developer trials remain limited. Daniel Dewhurst, an AI engineer who ran early tests on V4 after its release, said the preview appears significant but advised caution in interpreting benchmark claims until more independent evaluations and real-world usage data are available. He also noted that the preview highlights how open models that users can run themselves are closing the gap with closed-source alternatives in several dimensions, especially cost, extended context windows, and coding performance.
DeepSeek’s announcement included technical claims about V4’s context capabilities. The company said the model can process in excess of one million tokens, placing it on par with the long-context windows reported for other leading models. According to DeepSeek, V4 can handle such extended contexts while employing a fraction of the computation otherwise required by comparable systems.
The company’s broader corporate trajectory is part of the story. DeepSeek is owned by High-Flyer Capital Management and, according to a recent report cited by market watchers, seeks to raise capital at a valuation in excess of $20 billion. That report also said Alibaba and Tencent were in discussions to take stakes in the firm.
Despite restrictions in some jurisdictions, DeepSeek’s models continue to be among the most used on international platforms that host open-source AI models. The V4 preview, whether ultimately seen as a landmark or incremental step, is likely to shape discussions about compute supply chains, cross-border technology dependencies, and the competitive dynamics among Chinese AI developers and chip manufacturers.
Key points
- DeepSeek released a preview of V4, including a Pro variant optimized for Huawei Ascend chips, signalling a reduced reliance on Nvidia hardware.
- V4 ranks ahead of other open-source models on world-knowledge benchmarks, trailing only Google’s Gemini-Pro-3.1, and supports very long context windows exceeding one million tokens.
- Markets reacted: Chinese chipmakers rallied on expectations for homegrown chip adoption, while some AI rivals’ shares fell following the V4 announcement.
Risks and uncertainties
- Regulatory and geopolitical friction - U.S.-China tensions and recent allegations of intellectual property theft create uncertainty for cross-border technology partnerships and component shipments, potentially affecting chipmakers and cloud providers.
- Benchmark and real-world performance - Independent evaluations and broader developer testing are still limited; real-world efficacy and generalizability of preview benchmark claims remain to be validated, which affects enterprise and developer adoption.
- Market and competitive response - DeepSeek’s position in China has been eroded by many domestic competitors, and investor sentiment may shift as rivals and hardware providers react to V4’s capabilities and deployment choices.