Artificial intelligence is no longer the monopoly of American giants. With DeepSeek V3, China demonstrates its ability to develop high-performance AI models that, above all, are accessible to all.
DeepSeek, a Chinese company backed by High-Flyer Capital Management, has unveiled its latest AI LLM model: DeepSeek V3. And the first results are impressive.
To go further
What is an LLM? How do the engines of ChatGPT, Gemini and others work?
Big raw power
With its 671 billion parameters (the “neurons” of AI), DeepSeek V3 literally crushes the competition in terms of raw power. To put this figure into perspective, it is 1.6 times more than Meta’s Llama 3.1, previously considered a benchmark in the field.
This power translates into exceptional performance in many areas: coding, translation, writing, etc. The model particularly excels in programming tests on Codeforces, where it even outperforms OpenAI’s GPT-4o.
But what makes DeepSeek V3 truly remarkable is its value for money. The company claims to have spent only $5.5 million on its development, a pittance compared to the hundreds of millions OpenAI invested for GPT-4.
Some limitations
However, the model has some limitations, particularly in terms of hardware requirements.
Its imposing size requires substantial infrastructure to operate efficiently.
Even more problematic, the model reflects certain Chinese political constraints. Subject to Chinese government regulation, DeepSeek V3 carefully avoids certain sensitive topics.
Despite these restrictions, the impact of DeepSeek V3 is there. With a particularly competitive cost of use via API ($0.27/million tokens input, $1.10/million output), it represents a serious alternative to more expensive Western models.
How to use DeepSeek V3?
For those interested in testing with this new AI model, there are several ways to access DeepSeek V3.
The easiest method is to use the official web interface available at chat.deepseek.com. This platform allows you to interact directly with the model and even includes an Internet search function to obtain answers in real time. This is the ideal solution for beginners or those who want to quickly test the model’s capabilities.
For developers and more technical users, DeepSeek V3 is available on Hugging Face, the leading platform for AI models.
The interest of DeepSeek V3 lies in its permissive, open-source license, which authorizes its use for most applications, including commercial ones. Developers can therefore not only use the template, but also modify it to adapt it to their specific needs.
DeepSeek released the model on GitHub and a detailed technical document which describes his abilities.
Source: www.frandroid.com
