3At the beginning of 2025, the artificial intelligence and neural networks industry was struck by the unexpected appearance of the DeepSeek chatbot from a Chinese company of the same name.
A few months ago, the leading position in the field of innovative developments and AI remained with the American research organisation OpenAI, but now the situation is rapidly changing. DeepSeek's language model provides users with the same features as the well-known ChatGPT, and it is completely free of charge. Let's find out in detail what other advantages DeepSeek has, what the features of the new neural network are, and whether there are any differences from OpenAI models.
Who came up with DeepSeek, and how much did it cost?
Until January 2025, few people had heard of DeepSeek — the Chinese company's developments were of interest only to real industry enthusiasts. Although DeepSeek, an AI research company, was founded in Hangzhou back in 2023. And one of its founders, Liang Wenfeng, has been innovating and building experimental models since as early as 2015. Together with other Zhejiang University graduates, he founded the hedge fund High-Flyer, which used machine learning to research the stock market and buy and sell stocks. High-Flyer AI, a subsidiary solely focused on neural network development, soon emerged. It was renamed DeepSeek in 2023. The same organisation also funded the creation of DeepSeek-R1, a large open-source language model that can be applied to all areas of activity, not just the financial sector.
Even before the acclaimed DeepSeek-R1, the company developed the DeepSeek Coder, which is designed exclusively for writing and refining code. Among other DeepSeek developments, the DeepSeek-Math neural network for solving mathematical problems is also popular. But with the appearance of DeepSeek-R1 these neural nets are no longer in such demand, as the new model includes all the features of DeepSeek-Coder, DeepSeek-Math and other features of the same ChatGPT from OpenAI. At the same time, it took only 55 days and no more than $5.6 million to develop the new version of DeepSeek. According to preliminary estimates, it cost $100 million to create and train the GPT-4 model, which is a direct competitor of DeepSeek. You have to agree, the difference is impressive. While the researchers are trying to figure out how this is possible, let's look at the new neural network's functions.
DeepSeek feature set
In fact, the capabilities of the new Chinese AI are almost the same as those of its main competitor, ChatGPT. In other words, DeepSeek is able to work with texts in a variety of formats, for example, generate them from scratch, edit them, proofread and correct errors, and supplement them with information while maintaining a uniform style. The neural network will easily write an article, essay, letter, or even a poem according to the given criteria.
The new AI is also capable of quickly searching for data, cross-checking it and providing only the data that corresponds to reality. DeepSeek is also able to work with big data — to analyse it, structure and systematise it according to several classifications, and interpret the information.
In addition, you can just chat with the neural network — DeepSeek can simplify complicated topics, help you learn a foreign language, quickly find information, and also give advice and useful recommendations, for example, about what film you will like, what book you should read, or what artist you should listen to. Surprisingly, you can use DeepSeek to plan trips and all kinds of travelling experiences without any problems. All you have to do is describe your wishes and specify your budget, and the neural network will pick out a few options that you'll love!
The AI can also be used to write codes from scratch, refine them, and correct errors and inaccuracies. If you ask DeepSeek for more details about the error, the neural network will explain why something does not work in this case, how else it can be corrected, and how to avoid similar errors in the future.
Another important feature is DeepThink. It suggests that the neural network thinks in stages, i.e. it first receives and processes and analyses the incoming request, then it thinks through its answer, develops its structure, finds and corrects flaws, and only then gives the user the result. Interestingly, the decoding of the chatbot's “thought” process is available and can be used if DeepSeek makes a mistake. Nevertheless, the DeepThink configuration significantly improves the quality of the responses, while the time for the neural network to search, process and structure the information remains the same.
The chatbot works not only with users' text requests and documents. Although this mode is quite limited, you can upload documents and images to enable information recognition. For example, if an instruction in a foreign language causes you difficulties, you can easily upload it to the neural network and get answers to all your questions.
One of the few features not available in DeepSeek-R1 is picture generation. The neural network quickly comes up with ideas and describes images, but it can't create them because it is considered primarily a language model. So, it can write an exact prompt for another AI to generate a visual sequence for you. You can do this, for example, in Janus-Pro, which is another model from DeepSeek designed specifically for fast image recognition and generation.
Getting started with DeepSeek
Interestingly, all DeepSeek developments can be found in the public domain, as the model is based on open-source software. This means that all researchers, industry enthusiasts and just hobbyists can clearly see how the new neural network works. In addition, all of the above listed features are absolutely free of charge.
You can download DeepSeek to your phone or just use the browser version. The neural network appeared in the App Store and Google Play apps in mid-January and immediately created a stir, becoming a top download in the US and then in other countries.
To access the web version, all you need to do is go to the official website, click Start and log in (it's completely free and available in all countries worldwide). Don't worry about the site being available in only two languages, English and Chinese, either. The neural network itself understands multiple languages — when asked about this, DeepSeek replies that it understands and can communicate in a multitude of languages, including but not limited to: English, Spanish, French, German, Russian, Japanese, Arabic, Portuguese, Italian and many more.
The interface of DeepSeek is pretty simple and similar to other such chatbots — most of the screen is occupied by a dialogue box and a field for entering prompts. There are also two additional buttons there — DeepThink and Search. By pressing the former, you will get not only information on the question you asked, but also a full transcript of the AI's thinking before giving you the answer. In other words, you will learn in detail about the process of its considerations. The second button—Search—involves using large sources of information, analysing and comparing them. The quality of DeepSeek's answers is particularly high, but this can take about 10–20 seconds.
You can create several chats for different tasks in DeepSeek and seamlessly navigate between them. For convenience, you can rename them and delete the ones you no longer need.
In addition to the web version and the app, DeepSeek can also be used locally through special frameworks such as Ollama for running and managing large language models. This will ensure that the selected language model is loaded and integrated into the system, and can be accessed via API. However, this will require special equipment, as a neural network with a huge number of parameters and settings simply cannot be run on an ordinary laptop or computer. You will need a lot of RAM and a powerful graphics card.
How DeepSeek has impacted the industry
The popularity and instant success of the Chinese neural network provoked a fall in the shares of American companies in the field of development. For example, NVIDIA, the world's leading manufacturer of graphics processors, was hit hard. The technology company lost aroundabout $500 billion in capitalisation. Amid the success of the Chinese startup, its shares dropped in price by more than 15 per cent. Experts estimate the total reduction in the capitalisation of US technology giants such as NVIDIA at more than a trillion dollars.
In Europe, in turn, shares of Siemens Energy AG, a supplier of equipment up for the IT industry and AI development, plunged substantially — by almost 20 per cent. The share prices of ASML, ASM International, STMicroelectronics and others also declined.
Nevertheless, the emergence of DeepSeek has been called a real breakthrough in the industry. Especially considering how much it cost to train the new model. This has been possible in the face of US sanctions that prohibit the sale of chips to China, prompting companies such as DeepSeek, for instance, to seek and successfully find more efficient and lower-cost solutions. However, the low cost of developing AI like DeepSeek's makes one question the effectiveness of multi-billion dollar investments in the entire innovation industry. Thus, DeepSeek, using more affordable and efficient solutions, uses unique approaches to data processing, requires significantly less computational resources, and rightfully becomes a threat to OpenAI models. Leading experts believe that open models such as DeepSeek-R1 contribute to the active development of the industry and make neural networks more accessible to ordinary users.
Share this with your friends!
Be the first to comment
Please log in to comment