The newly released DeepSeek-R1, which rapidly gained global popularity, is not the company’s first release. Back in December 2024, they launched DeepSeek-V3, a GPT4o (the core of the previous ChatGPT version) competitor that could generate text and answer questions. The latest release, though unveiled just a month later, outperforms an analogous model, o1, by OpenAI: DeepSeek-R1 can “think” like a human, solve complex tasks in math, chemistry, and programming, create logical chains, and test its conclusions. At the same time, the new model is open-source, much cheaper, and  free.

Key DeepSeek features

• The latest AI models, such as GPT o1 (ChatGPT), require a lot of graphics card power, while DeepSeek can run on a smaller number of these devices and demonstrate results comparable to the best models. 

According to Anton Kuznetsov, the head of the Institute of Applied Computer Science, such performance is enabled by an improved architecture and a limitation on the number of parameters. DeepSeek-R1 can be launched on a single powerful computer, while its quality is largely comparable to that of larger models – with answers to questions of various difficulty levels, from general knowledge to coding.

• Since DeepSeek is an open-source product, companies around the world can train it on their hardware without the need to rent additional cloud resources or support complex infrastructure.

“With an open-source approach to research and products, it’s possible to develop and promote your own solution. This is demonstrated by both international and Russian companies, including Yandex. This approach makes it possible to involve specialists and researchers from other countries in your projects for free,” shares Vladislav Gorbunov, the head of AI at the Institute of Applied Computer Science.

• Another feature of DeepSeek is its familiarity with Chinese culture – the model understands dialects, idioms, historical references, and modern social trends, which isn’t something accomplished by its Western counterparts. Moreover, DeepSeek complies with Chinese regulations, meaning that it can be safely used by government and corporate services, such as for automation of support services, analysis of local markets, development educational materials, or processing of data within state initiatives. For instance, the model helps businesses optimize their Taobao logistics and assists students learning classical Chinese texts.

Is it true that DeepSeek is more powerful than ChatGPT in every way? 

DeepSeek can perform similarly to ChatGPT, while being significantly cheaper and less computationally demanding. It took two months and $5.5 million to get DeepSeek to its current level, while the training of GPT-4 came down to over $100 million.

“DeepSeek was founded in 2015 and started purchasing graphics cards in 2021, just as GPT-3, a relatively good text generator without logical thinking, was released. Thus, the startup team had set their goals long ago – they didn’t get into LLMs yesterday. At the same time, we can’t say that DeepSeek is significantly better than ChatGPT. DeepSeek R1 is comparable to ChatGPT o1, while DeepSeek-V3 performs on par with ChatGPT 4o,” adds Vladislav Gorbunov.

Among the key factors that have contributed to DeepSeek’s success are its low computational demands and open source. By late January, DeepSeek’s app surpassed ChatGPT on the App Store in the US and then in Russia. Among the Russian audience, the model gained popularity partly due to limited access to its analogs, including ChatGPT, Gemini, and others. DeepSeek is available for free for iOS, Android, and on the company’s website.

Is DeepSeek a breakthrough? 

“DeepSeek is a new development in the field of LLMs: whereas earlier, developers were aiming for the highest performance regardless of the required resources, now comes a time when everyone will be trying to maintain high quality while cutting down on expenses. In the nearest future, we can expect models to become even more optimized and, potentially, we may even see models that will function on a smartphone, offline, and offer results comparable in quality to LLMs,” says Anton Kuznetsov.

How will DeepSeek impact the Russian IT industry and business? 

In Russia, there are similar solutions: YandexGPT and GigaChat by Sberbank. Even though they are built on the classic LLM architecture, their results when generating Russian texts are in some cases better than those of their counterparts. DeepSeek’s experience demonstrates that even without major financial contributions and giant powerhouses, it’s possible to create innovations that can outperform global leaders.

“With enough resources, it’s possible to replicate DeepSeek’s success in Russia – all the involved research and approaches are open-access. However, efforts should be focused on preparing a dataset fitted to our culture, following in the footsteps of our Chinese colleagues, or, alternatively, a dataset for a specific field with all its specific jargon (such as in medicine or law). DeepSeek, like many current AI projects, is based on open-source research (transformer architectures, RLHF) that can be adapted for local tasks. Our experts can use the same approach, without wasting any time on reinventing the wheel,” concludes Vlad Gorbunov.