DeepSeek V4 AI Released, Becoming China's Exit from Dependence on Nvidia

China’s efforts to reduce dependence on artificial intelligence (AI) chips from the United States are becoming increasingly evident. The Chinese AI startup DeepSeek has released a preview of its latest AI model, optimised to run on Huawei’s Ascend chips. The AI model, named DeepSeek V4, serves as an “exit” and a strong signal that China’s AI ecosystem is shifting away from reliance on Nvidia’s GPU systems. As is well known, most global AI models, including those in China, have been developed using Nvidia’s CUDA ecosystem for training and inference processes. Now, DeepSeek is breaking this pattern of dependence, as the V4 model is claimed to have completed inference adaptation on Huawei’s Ascend platform. This indicates that implementation on domestically produced chips by Chinese AI companies is transitioning from the testing phase to practical application. Notably, DeepSeek has for the first time placed Nvidia GPUs and Huawei’s Ascend NPUs within a single hardware validation framework. The company states that the fine-grained expert parallelism scheme they developed has been verified to run on both platforms. He Hui, Director of Semiconductor Research at consultancy firm Omdia, described this step as falling into the category of “major development”, particularly for China’s AI industry. “Huawei’s Ascend chip is the best domestic alternative to Nvidia, and support for DeepSeek V4 shows that top Chinese AI models can now run on Chinese-made hardware,” said Hui, quoted by KompasTekno from Reuters. In terms of performance, DeepSeek claims that the V4 model can process more than one million tokens, equivalent to the context capacity of models like OpenAI’s GPT-5.4 and Anthropic’s Claude Opus 4.6. Tokens themselves are units for counting the number of characters or words processed or generated by an AI model. One token can mean one character or one word in a sentence, depending on the complex system calculation. The DeepSeek V4 AI model also claims to offer several improvements in agent capabilities, world knowledge, and reasoning performance, which are key benchmarks in evaluating next-generation AI. Meanwhile, the Pro version is claimed to outperform other open-source models, particularly in world knowledge benchmarks, and only trails Google’s closed model Gemini-Pro-3.1.