DeepSeek V4: The Chinese AI Challenging American Giants

⚡

Key Takeaways

1DeepSeek V4, developed by a Chinese company, offers open models that compete with those from Google and OpenAI.

2The DeepSeek-V4-Pro model stands out for its performance in mathematics, science, and coding, ranking sixth according to Artificial Analysis.

3The models are available in beta on DeepSeek Chat and in Open Weights on Hugging Face, making it easier for developers to access.

💡Why it matters — DeepSeek V4 could transform the AI market by providing open alternatives to American solutions, influencing innovation and overall competitiveness.

DeepSeek V4: A New Competitor in the AI Market

Last year, the release of the DeepSeek R1 model already shook the American giants of artificial intelligence with its ability to produce advanced reasoning while conserving resources. This week, the Chinese company DeepSeek launched its new generation of AI models, named DeepSeek V4. These models, while open, present themselves as true alternatives to the technologies of industry leaders like Google, OpenAI, and Anthropic.

DeepSeek-V4-Pro: High-Level Performance

The DeepSeek-V4-Pro model is particularly noteworthy for its advanced capabilities in areas such as mathematics, science, and coding. DeepSeek claims that this model is now the best open AI for these disciplines. In terms of knowledge and agentic skills, which include the ability to perform concrete tasks, DeepSeek-V4-Pro approaches the performance of technologies offered by American giants.

Enhanced Agentic Capabilities: This model is at the forefront of open-source in agentic coding benchmarks.
Rich World Knowledge: It surpasses all current open models, being outperformed only by Gemini-3.1-Pro.
World-Class Reasoning: It excels in mathematics, science, and coding, competing with the best.

In the Artificial Analysis ranking, which uses GDPval-AA to evaluate AI skills in professional tasks, DeepSeek-V4-Pro ranks sixth. Its score is close to that of Opus 4.6 from Anthropic, demonstrating its competitiveness in the market.

DeepSeek-V4-Flash: Efficiency and Resource Savings

On the other hand, DeepSeek-V4-Flash is a smaller model that allows for resource savings. Despite its reduced size, DeepSeek claims that its reasoning capabilities are comparable to those of the Pro version, which could make it highly competitive, especially for applications requiring less computational power.

Testing DeepSeek V4: An Accessible Opportunity

Although it is not as powerful as the latest technologies from OpenAI, Google, and Anthropic, DeepSeek V4 demonstrates that China can produce artificial intelligence models whose performance is close to that of American competitors. Moreover, the DeepSeek V4 models are open, making them accessible to a wide audience.

For now, this is a pre-release version. Nevertheless, DeepSeek indicates that it is already possible to test these new models on DeepSeek Chat. The "Open Weights" versions are also available on the Hugging Face platform, allowing developers to explore and create new features based on these models. DeepSeek's API also offers the possibility to develop applications integrating these new technologies.

A Potential Impact on the AI Market

Even though American AI models remain the best, DeepSeek has just launched a viable alternative to them, having already caused a stir in the stock market after the release of DeepSeek R1 in 2025. Furthermore, DeepSeek V4 also challenges closed models, as it is available in Open Weights. The Chinese company DeepSeek has just launched the DeepSeek V4 model family. This is the best family of open models, and its performance is close to that of American closed models. It is already possible to test these new DeepSeek models on DeepSeek Chat.