What Is DeepSeek-V4? China’s Latest AI Model That Ditches Nvidia for Huawei Ascend Chips

Chinese AI startup DeepSeek has unveiled its latest artificial intelligence model, DeepSeek-V4, in what could become a major milestone for China's push toward AI and semiconductor self-reliance.

Unlike many competing AI systems that rely heavily on Nvidia hardware, DeepSeek-V4 has been specifically optimized for Huawei Ascend AI chips, reducing dependence on foreign processors amid ongoing US export restrictions. The launch highlights China's growing efforts to build a fully domestic AI ecosystem capable of competing with major US firms such as OpenAI and Google.

DeepSeek-V4 Comes in Two Versions

DeepSeek-V4 is available in two variants - V4-Pro and V4-Flash. The Pro model is designed for high-performance workloads and advanced AI tasks, while the Flash variant targets lighter workloads with lower operating costs.

According to DeepSeek, the V4-Pro model outperforms most open-source AI systems in world knowledge benchmarks and trails only Google's Gemini-Pro 3.1 among leading AI models. Reports also suggest the model competes with GPT and Gemini systems in coding, reasoning, and agentic AI capabilities.

The AI system is also designed to work with agent frameworks such as Claude Code and OpenClaw, enabling automated workflows for coding and other advanced applications with minimal manual input from developers.

Massive 1-Million-Token Context Window

One of the biggest highlights of DeepSeek-V4 is its 1-million-token context window. This allows the AI model to process extremely large documents, maintain longer conversations, and handle complex coding tasks more efficiently than many previous-generation models.

DeepSeek also claims the V4 architecture improves computing efficiency and reduces memory requirements compared to earlier DeepSeek models, potentially lowering operational costs for enterprises and developers.

Why DeepSeek Is Using Huawei Chips Instead of Nvidia

A major part of DeepSeek-V4's launch revolves around its deep integration with Huawei Ascend AI hardware. Huawei confirmed that its Ascend 950 chips were used during portions of the model's training process, while the company's Ascend supernode infrastructure will fully support deployment of the V4 series.

The move comes as Chinese AI firms continue facing restrictions on access to advanced Nvidia processors due to US export controls. By optimizing its models for Huawei's domestic AI chips, DeepSeek is attempting to reduce reliance on foreign semiconductor technology.

DeepSeek also expects costs for the Pro version to decline further as Huawei Ascend supernodes expand across China. The company believes cheaper access to local AI compute infrastructure could make advanced AI services more accessible to developers, researchers, and businesses nationwide.

DeepSeek's Growing Role in the AI Race

DeepSeek previously gained international attention with its low-cost DeepSeek-R1 model earlier in 2025, which disrupted the AI market by delivering competitive performance at significantly lower training costs.

With DeepSeek-V4, the company is now positioning itself as one of China's strongest challengers in the global AI race. Industry analysts see the DeepSeek-Huawei partnership as a direct challenge to Nvidia's dominance in AI hardware and an important step toward China's broader AI independence goals.

With advanced reasoning capabilities, a massive context window, and lower-cost deployment ambitions, DeepSeek-V4 could become one of China's most important domestic AI platforms in the coming years.