Chinese AI startup DeepSeek has quietly released its new model, DeepSeek V3.1, featuring 685 billion parameters. The company, with the support of High-Flyer Capital Management, has uploaded this model directly to the Hugging Face platform without a major launch.
DeepSeek V3.1 Unveiled
While this move may seem modest at first glance, quick tests have revealed that the model can easily compete with the most advanced systems from giants like OpenAI and Anthropic. The model’s open source nature makes it accessible worldwide, free from geographical or political restrictions.
The arrival of DeepSeek V3.1 heralds not only a technological advancement but also a fundamental shift in the development, deployment, and control of artificial intelligence. This pushes the technological rivalry between the US and China to a new level.
With an Aider coding score of 71.6 percent, DeepSeek V3.1 directly challenges its American competitors. It can process up to 128,000 context tokens, equivalent to a roughly 400-page book, and its response times are faster than previous “thinking” models.
The system supports a variety of precision formats, from BF16 to the experimental FP8, allowing developers to optimize for their hardware. The company says this model is optimized for Chinese-made chips and offers higher processing performance.
The most striking aspect of the model lies in its approach, which they call a “hybrid architecture.” V3.1 combines chatbot, logical inference, and coding capabilities into a single system, delivering consistent, high-performance results. According to AI researcher Andrew Christianson, DeepSeek V3.1 outperforms Claude Opus 4 by 1 percent while reducing cost by a factor of 68.
The model’s architecture also includes search tokens for real-time web integration and think tokens for internal reasoning. These features demonstrate that it has overcome fundamental problems previously encountered in hybrid systems.
DeepSeek V3.1 also delivers impressive efficiency. At a cost of approximately $1.01 for a full coding task, it delivers results comparable to competing systems costing around $70. For companies requiring thousands of AI interactions daily, this cost difference translates into potential savings of millions of dollars.
By maintaining its open source nature, DeepSeek directly challenges the business models of US-based companies. While US companies protect their models with high licensing fees and usage restrictions, DeepSeek offers free access to similar capabilities. This further intensifies competition in the AI market.
