DeepSeek has launched preview versions of its new V4 model, strengthening its position in the escalating artificial intelligence race with the United States. The release follows months of anticipation and targets direct competition with systems such as ChatGPT, Claude, and Gemini.
The company says V4 marks a significant leap in knowledge and reasoning capabilities, alongside enhanced agent-like features that allow the model to carry out complex tasks independently. It also confirmed partial reliance on Huawei chips in an effort to reduce dependence on US suppliers.
Details
The V4 release builds on the earlier V3 model launched in 2024, while the company’s real breakthrough came with the R1 model in early 2025, which surprised markets with its efficiency and low cost.
The new version includes two main variants:
- Pro version for advanced performance
- Flash version for speed and efficiency
The Pro Max variant reportedly outperforms some US models in reasoning benchmarks, while approaching — but not surpassing — the latest releases. The model also features a context window of up to one million tokens, compared to 128,000 in the previous version.
Meanwhile, US companies including OpenAI and Anthropic continue to accuse DeepSeek of leveraging distilled techniques derived from their models. US officials have echoed these claims, accusing Chinese firms of exploiting American innovation—allegations Beijing rejects as unjustified attempts to curb its progress.
What’s Next?
The real test will come from independent evaluations and real-world performance, as competition intensifies and the AI landscape further splits between China and the United States.
