DeepSeek V4, a More Efficient and Powerful AI than ChatGPT

DeepSeek, the Chinese artificial intelligence company, has made a splash with the launch of its V4 model family. Unlike a year ago, when it mainly competed with OpenAI and Google, it now faces other advanced AIs like Kimi K 2.6. The V4 series includes two models based on the Mixture-of-Experts (MoE) architecture, designed to handle long contexts more efficiently.

The first, DeepSeek-V4-Pro, has 1.6 trillion parameters but activates only 49 billion during inference. The second, DeepSeek-V4-Flash, works with 284 billion parameters and activates 13 billion per query. Internal tests show that the Pro’s reasoning mode ranks it among the best open-source models for various tasks, even matching GPT-5.4 in competitive programming.

Compared to giants like Gemini 3.1-Pro or Claude Opus 4.6, DeepSeek V4 still trails in general knowledge and reasoning. However, it excels in retrieving information from long texts. Its key innovation lies in combining compressed sparse attention and highly compressed attention, significantly reducing computational resource needs. This makes DeepSeek more efficient than previous versions and other competitors, while maintaining high performance in complex tasks.

The improved architecture allows DeepSeek V4 to operate with only 27% of the computational operations of its predecessor, making it highly attractive for resource-demanding applications. The tech community is watching closely to see how this AI can reshape the landscape, offering a powerful and efficient alternative to market giants.