Deepseek V3: Chinese open source AI with censorship

The Chinese AI model Deepseek V3 is said to be able to keep up with the best models from OpenAI and Google in numerous benchmarks. The model is open source – and is said to outperform other freely available models. According to the provider, the development and training of Deepseek V3 cost significantly less than other large models. The problem is that Deepseek V3, like other models from China, censors content and answers. This can be circumvented with tricks, but you have to know when this is necessary – i.e. what a censored answer was that needs to be circumvented.

Deepseek V3 is basically a mixture of experts and comes with a chatbot that you can already try out. Deepseek V3 can also be integrated via API. Mixture-of-experts means that only individual experts who are suitable for the task are addressed when responding. The model is based on 671 billion parameters. According to the provider on Github, only 2.788 million GPU hours on Nvidia's H800 GPUs were required for training. This already includes fine-tuning and further rounds of reinforcement learning. A model learns through confirmation. The predecessor was called Deepseek R1 and specialized in reasoning. Similar to OpenAI's o1, the AI responds in loops, i.e. it rethinks its own response.

According to AI expert Andrej Karpathy, Meta's free model Llama 3 required around 30.8 million GPU hours for 405 billion parameters. He describes the budget for Deepseek V3 as a "joke" at around six million US dollars. This sum has not been confirmed by Deepseek. It is also unclear where the training data comes from. There are speculations that it is based on responses from ChatGPT. Deepseek V3 occasionally responds by claiming to be ChatGPT itself.

Good benchmarks and censored responses

Deepseek V3 performs almost as well or even better than other free models in numerous benchmarks. In the Math-500 math test, Deepseek V3 achieves 90.2 percent while GPT-4o only achieves 74.6 percent. Unsurprisingly, the Chinese model is also particularly good at speaking and understanding Chinese. However, its ability to speak German, for example, is not known. The choice of training material is also decisive for language skills. If a model has been trained primarily with English texts, as was initially the case with the major providers, they will struggle with other languages – and possibly other cultures –.

Deepseek V3 has published the benchmark results on its own website. They are therefore not independently verified figures. In addition, such tests are not very meaningful in terms of how useful and helpful models are in everyday life.

It is also problematic that Deepseek V3 answers some questions in the interests of the Chinese government. For example, there is no criticism of the government. Events on Tiananmen Square are also concealed by the AI. Protests by a democracy movement were bloodily ended there in 1989. The usual AI tricks are used to get the model to write about the massacre. But if you don't know that something is being concealed, it's difficult to use tricks because you don't even know that something is missing.

Nevertheless, Yann LeCun, AI expert and head of Meta, also says that Deepseek V3 is "excellent".

(emw)