DeepSeek unveils V4 AI model with significant improvements
DeepSeek has unveiled the V4 AI model, with significant gains in efficiency and performance, narrowing the gap with today's leading models.
DeepSeek, the Chinese artificial intelligence startup, has recently previewed its new AI model, DeepSeek V4. This update is a significant step up from the previous V3.2 model, thanks to architectural improvements that increase its efficiency and performance. According to the company, V4 has nearly closed the gap with today's leading models, both open and closed, on reasoning benchmarks.
Key features of DeepSeek V4
DeepSeek V4 comes in two versions: V4 Pro and V4 Flash. Both use a "mixture-of-experts" architecture with a context window of 1 million tokens, allowing them to process large documents or codebases. The Pro model has 1.6 trillion parameters, with 49 billion active, while the Flash version has 284 billion parameters, of which 13 billion are active. These numbers make V4 one of the most powerful open-source models currently available.
Performance on reasoning benchmarks
DeepSeek claims that the V4 Pro Max model outperforms competing open-source models on reasoning benchmarks, surpassing OpenAI's GPT-5.2 and Google's Gemini 3.0 Pro on some tasks. However, in some knowledge tests V4 lags slightly behind leading models such as GPT-5.4 and Gemini 3.1 Pro, indicating a development gap of about 3-6 months compared with the most advanced models.
"Agentic" capabilities and practical applications
V4 Pro is designed to carry out complex tasks autonomously, with "agentic" capabilities that let it handle complex workflows without human intervention. In internal tests, V4 Pro outperformed Claude Sonnet 4.5 and comes close to the level of Claude Opus 4.5, positioning itself as a powerful solution for applications that require autonomy and advanced reasoning.
Implications for the market and the AI industry
The release of DeepSeek V4 is a significant challenge for industry leaders such as OpenAI and Google, offering an open-source model that is highly competitive in both performance and cost. With a context window of 1 million tokens and advanced reasoning and autonomy capabilities, V4 is set to significantly influence the development of AI applications across many sectors, from programming to managing complex workflows.
Conclusion
- Architectural innovation: DeepSeek V4 introduces significant improvements that increase efficiency and performance compared with previous models.
- Competition in the AI market: With performance comparable to leading models, V4 is a challenge for industry leaders such as OpenAI and Google.
- Advanced practical applications: V4's "agentic" capabilities make it suitable for applications that require autonomy and advanced reasoning.
- Implications for future development: V4 could accelerate the adoption of advanced AI models across many sectors, influencing the development of applications and services based on artificial intelligence.