China's free open-source AI DeepSeek is a serious threat to OpenAI's ChatGPT and other AI models

From November 2023 onwards, the Chinese company DeepSeek has been releasing AI models on an open-source basis. Thanks to the MIT license, anyone can use and adapt the model for their own individual purposes. This makes the models transparent and versatile.

Additionally, they enable a collaborative development and save money along the way. Users can view and understand the code in order to see how the model works. They can adapt the model to their specific needs and use it for various applications. Therefore, DeepSeek is committed to open source and thereby adds  innovation and competition to the AI field.

This company emerged from Fire-Flyer, which is the deep-learning branch of a Chinese hedge fund named High-Flyer. The initial goal was to better understand, interpret, and predict financial data in the stock market. Since the spin-off of DeepSeek in 2023, the company has been focusing entirely on LLMs, which are AI models capable of generating text.

The company seems to have achieved a major breakthrough with the two newest members of the DeepSeek AI family. According to widely used AI benchmarks, DeepSeek-V3, DeepSeek-R1, and DeepSeek-R1-Zero often outperform competitors from Meta, OpenAI, and Google in their respective fields. As an online service, they are also significantly cheaper than ChatGPT, for example.

This aggressive pricing strategy could have an impact on all prices on the AI market and make advanced AI tools accessible to a wider audience. The company can afford this because it wants to spend significantly less money on training its AI models than other companies do. To be more precise, this can be achieved through more efficient training programs and lots of automation.

On the other hand, DeepSeek-R1 and DeepSeek-R1-Zero are reasoning models. This means that they first develop a plan on how to best answer a question and then work in small steps. This improves the accuracy of results while requiring less computing power. However, this approach increases the demand of storage space.

As an open-source AI, DeepSeek can run directly on the end user's computer. The required application data can be accessed for free, as the models can be downloaded free of charge on Hugging Face. It's even easier with programs like LM Studio, which can automatically download and install the entire code of the application.

This means that there should be no issues with data security and privacy. Prompts, data, and answers do not leave the computer. In addition, the model can be used offline. High-performance hardware is not required here, but a lot of memory and storage is. For instance, DeepSeek-R1-Distill-Qwen-32B requires roughly 20GB of storage space on the hard drive.

According to DeepSeek V3, the AI ​​works with numerous languages. These include Chinese and Englis  but also German, French, and Spanish. In a short chat, the various languages provided satisfactory answers.

All that remains are questions and doubts about China’s censorship. DeepSeek-R1 contains censorship for certain politically sensitive topics. Users who try to ask about certain historical events either receive no answer or a "revised" answer. For example, you don't necessarily have to ask the AI about what happened at Tiananmen Square on June 3rd and 4th, 1989.

That said, DeepSeek R1 does at least recognize the student protests and a military operation. But other AI models also skimp on answers to political questions. Google's Gemini outright refuses to answer questions might be related to politics. Therefore, (self-imposed) censorship can be found in various AIs.