DeepSeek V4 Launches on Huawei Chips as Open-Source AI Challenges GPT-5.5 and Gemini 3.1-Pro

DeepSeek has released preview versions of its latest AI model, DeepSeek V4, marking a significant update in its series of open-source large language models. The release comes after a period of relative quiet following the market impact of its earlier V3 and R1 models, which demonstrated that high-performance AI could be developed with relatively modest resources.

The V4 series includes two versions: DeepSeek-V4-Pro and DeepSeek-V4-Flash. The Pro variant features 1.6 trillion parameters and a context length of one million tokens, equivalent to approximately 750,000 words. DeepSeek states that the V4-Pro model delivers performance rivaling the world’s top closed-source models, while the V4-Flash offers a smaller, more cost-effective alternative.

A notable aspect of the V4 release is DeepSeek’s use of Huawei’s Ascend AI chips for training the model. This represents a strategic shift away from reliance on Nvidia GPUs, which are commonly used by AI companies worldwide for training large models. The move underscores efforts by Chinese firms to develop technological autonomy in critical AI infrastructure.

DeepSeek emphasizes the open-source nature of its V4 models, allowing developers globally to access, modify, and build upon the technology. This approach aligns with the company’s earlier strategy that contributed to its initial traction in the AI community. According to data referenced in the Android Headlines report, Chinese models from firms like DeepSeek and Alibaba now account for nearly a third of global AI usage.

The company highlights improvements in the V4 model’s “agentic” capabilities, referring to the ability to perform complex tasks and workflows autonomously. DeepSeek claims the V4-Pro excels in agentic coding and reasoning, outperforming several open-source models in generating computer code and managing multi-step software interactions. These enhancements aim to help human programmers reduce repetitive tasks and focus on higher-level aspects of development.

In benchmark comparisons, DeepSeek states that its V4 Pro Max version demonstrates superior performance in standard reasoning benchmarks relative to OpenAI’s GPT-5.2 and Google’s Gemini 3.0-Pro models. However, the company acknowledges that the V4 model falls “marginally” short of OpenAI’s GPT-5.4 and Gemini 3.1-Pro in certain evaluations.

The release has drawn attention in financial markets, with shares in Semiconductor Manufacturing International Corp. (SMIC), which produces Huawei’s Ascend processors, rising 10% in Hong Kong trading following the announcement. Conversely, shares in two of DeepSeek’s competitors, MiniMax and Knowledge Atlas, declined by more than 9%.

DeepSeek faces ongoing scrutiny from Western technology firms and governments. OpenAI and Anthropic have accused the startup of “distillation,” a process involving imitating the behavior of other models by analyzing their outputs. Several governments have restricted the use of DeepSeek’s platform due to data privacy concerns.

Some analysts note that independent evaluations are still needed to verify all of DeepSeek’s performance claims. Nevertheless, the rapid pace of the company’s model releases indicates sustained competition in the global AI landscape, particularly between Chinese and U.S. Developers.