DeepSeek quietly releases new version of AI

This new model not only challenges the dominance of American giants but also reshapes the entire game with open source (Illustration: DS).

The model has 685 billion parameters and is released under an open source license. Initial testing results show that its performance is on par with proprietary systems from OpenAI or Anthropic, and opens up a new approach to developing and distributing advanced AI models.

Performance and cost

Shortly after its release on the Hugging Face platform, DeepSeek V3.1 scored 71.6% on the Aider programming test. Technically, the model is capable of handling contexts of up to 128,000 tokens (equivalent to a book of about 400 pages). The architecture is described as “hybrid”, integrating conversation, reasoning, and programming functions.

“DeepSeek V3.1 achieved 71.6% on Aider, about 1% higher than Claude Opus 4 while costing significantly less,” AI researcher Andrew Christianson said on social network X.

It is estimated that the average cost to complete a programming task using V3.1 is around $1.01, while equivalent systems can cost closer to $70.

Some community developers also noted the presence of four “special tokens” in the architecture, which allow the model to integrate real-time web search and handle complex inference processes.

This difference translates to millions of dollars in savings for businesses that handle thousands of AI interactions every day.

Game-changing advantage

The V3.1 release comes just weeks after OpenAI and Anthropic announced new models. By choosing an open-source approach, DeepSeek follows a different strategy than many U.S. companies that commercialize AI models through paid APIs.

Some argue that making advanced capabilities widely available could enable organizations and businesses to deploy more quickly, while also challenging business models based on technology monopolies.

The emergence of DeepSeek V3.1 shows that advanced AI is not only the domain of large corporations, but also smaller research groups.

Competition in this space is no longer limited to building the most powerful system, but also revolves around making technology accessible and affordable.

In that context, the question for technology companies is how to maintain commercial value when more and more open source models appear with competitive performance and cost.

Source: https://dantri.com.vn/cong-nghe/deepseek-am-tham-ra-phien-ban-ai-moi-20250820121017977.htm