Select Language:
DeepSeek, a Chinese startup that gained global recognition last year with its affordable AI model, unveiled a preview on Friday of a highly anticipated new model optimized for Huawei chip technology. This development highlights China’s increasing independence in AI innovation.
The Pro version of this new model surpasses other open-source AI benchmarks in global knowledge tests, coming only behind Google’s proprietary Gemini-Pro-3.1, according to DeepSeek.
Unlike its previous reliance on Nvidia’s AI chips, DeepSeek’s collaboration with Huawei on the V4 model marks a significant shift. Huawei stated that its chips were utilized during parts of the V4’s training.
He Hui, director of semiconductor research at consultancy Omdia, noted, “This is a major milestone for China’s AI sector. Huawei’s Ascend chips are currently the nation’s leading domestic alternative to Nvidia, and their support for DeepSeek V4 demonstrates that top Chinese AI models can now operate on Chinese-made hardware.”
Most prominent AI models are developed and run using Nvidia chips, but DeepSeek’s move toward Huawei’s technology underscores fears expressed by Nvidia CEO Jensen Huang. He worries that U.S. export restrictions and China’s push for self-reliance threaten Nvidia’s developer ecosystem in China.
Huang recently said on a podcast, “The day DeepSeek launches on Huawei hardware, that’s a serious concern for our country.”
The V4 model is currently the fastest to reach the top position on Hugging Face, a popular platform for sharing and deploying machine learning models, according to Lewis Tunstall, a machine learning engineer at Hugging Face.
It excels at comprehending extremely lengthy and complex texts and offers these capabilities at a much lower cost than competing leading models. However, it does have limitations, such as lacking support for multiple modalities like images and videos, Tunstall explained.
#### Huawei and DeepSeek’s Partnership
DeepSeek has faced criticism from the U.S. government and competitors across the West, who argue that its success is partly due to improper use of American technology. While the company has acknowledged using Nvidia chips, it has not clarified if these chips fell under export bans. It also stated it has not intentionally used synthetic data produced by OpenAI.
The launch comes shortly after the White House accused China of large-scale theft of U.S. AI research intellectual property. It also precedes a planned visit next month by U.S. President Donald Trump to Beijing to meet China’s leader Xi Jinping.
In January, the Trump administration approved the sale of Nvidia’s powerful H200 chips to China, but reports suggest that shipments have been delayed due to disagreements over sale terms in both countries.
Chinese chipmakers responded positively to the prospects of indigenous technology, with Huahong Semiconductor and SMIC rising by 15% and 10%, respectively.
Meanwhile, Nvidia’s stock also rose, buoyed by unexpected strong earnings forecasts from Intel, reaffirming confidence that the AI boom is far from reaching its peak.
#### Competition and Challenges
Many Western and some Asian governments have banned their agencies from using DeepSeek’s models, citing concerns over data privacy. Nevertheless, DeepSeek’s models remain heavily used on international open-source platforms.
In China, despite its rapid rise to champion status last year, its leadership has been challenged by a wave of domestic competitors. The release of V4 caused stock prices of rivals like Zhipu AI and MiniMax to tumble by 9%.
DeepSeek indicated that the V4 model is particularly suited for AI agents—a more sophisticated form of AI capable of executing complex tasks beyond simple chat functions, but which demands greater computing resources.
The ultimate success of V4 is still uncertain. Daniel Dewhurst, an AI engineer who evaluated the model post-release, remarked, “The initial impression is that V4 looks promising, but we should wait for independent reviews and more extensive real-world testing before drawing conclusions.”
He also pointed out that open models like V4 appear to be closing the performance gap with proprietary models, especially in areas like cost efficiency, processing long contexts, and coding tasks.
V4 can handle over a million tokens—comparable to the context windows of OpenAI’s GPT-5.4 and Anthropic’s Claude Opus 4.6—while requiring significantly less computational power. An entry-level Flash version of V4 is also available, allowing the company to gather user feedback and refine the model before a full launch. The timeline for finalizing the model remains unclear.
DeepSeek, owned by China’s High-Flyer Capital Management, is reportedly seeking funding at a valuation exceeding $20 billion. Reports indicate that major Chinese tech firms Alibaba and Tencent are in discussions to acquire stakes.
ChatGPT
Perplexity
Gemini AI
Grok AI
