On December 26, 2024, the Chinese company DeepSeek created a surprise by unveiling its new artificial intelligence model DeepSeek V3. This announcement marks a significant turning point in the AI race between China and the United States, particularly with its impressive performance and open source approach. Founded in 2023, DeepSeek has quickly established itself as a major player in the sector, developing several successful models in less than a year.
A technical and economic feat
DeepSeek V3 stands out with exceptional technical characteristics. With its 671 billion parameters (1.6 times more than Meta’s Llama 3.1), the model was trained on an impressive dataset of 14.8 trillion tokens. Performance is there: the model outperforms the competition in several areas, particularly in programming where it obtains better results than GPT-4o on the Codeforces platform.
The most remarkable aspect remains its development cost: only $5.5 million, a pittance compared to the hundreds of millions usually needed for this type of project. This feat was achieved thanks to the use of a data center equipped with Nvidia H800 GPUs for approximately two months, despite American restrictions on the export of these components to China. The execution speed is also impressive, with a processing capacity of 60 tokens per second, three times faster than the previous version.
An accessible alternative with Chinese specificities
DeepSeek’s open source approach allows for great flexibility of use. The model is accessible in several ways:
- The official web interface on chat.deepseek.comwhich even includes a built-in search engine
- The HuggingFace platform for developers
- An API with very competitive prices ($0.27/million tokens in, $1.10/million in output)
- The complete source code on GitHub with detailed technical documentation
DeepSeek V3 nevertheless has certain limitations. Its use requires a substantial infrastructure due to its large size. Additionally, being a Chinese company backed by the High-Flyer Capital Management fund, DeepSeek must comply with strict local regulations. The model thus avoids certain sensitive subjects and its responses “embody fundamental socialist values” according to the requirements of the Chinese internet regulator. Don’t expect to ask him if China is an authoritarian regime, he will give you a very vague answer.
This technological advance is part of a broader context of China’s rise in the field of AI. Other Chinese players like Alibaba (Qwen) or Tencent (HunyuanVideo) are also developing promising models, all in open source. This approach, different from that of the United States, allows researchers to move forward more quickly and access more data. The question now arises of the American reaction to this growing technological competition, particularly in the context of the upcoming elections and a possible Trump presidency, which could consider restrictive measures similar to those taken against Huawei.
