Software self-sufficiency: AI developers are increasingly using Open Source

Open Source, that is, the use of open models, increasingly helps to implement, train and improve neural networks faster. According to industry experts, active implementation of such solutions may begin in 2025. One of the growth drivers of this market is artificial intelligence (AI). The development of Open Source makes it possible to gain access to advanced technologies and maximize efficiency, even for small companies and organizations with limited IT budgets, experts say.

Why do AI developers need Open Source?

The use of open AI solutions in businesses of various scales is expanding, market participants say. Today, it is a full-fledged tool available both to those who have the necessary expertise in AI, and to those who are just starting to innovate in their business processes, they claim. Open Source helps to save significantly on development, for example, if developers get access to a ready—made model and retrain it to meet their own requirements.

There are other indisputable advantages.: An active community is usually formed around Open Source, technology security is improving, and the technology market as a whole is developing.

A bot to help you: businesses have started actively hiring AI directors

What are the responsibilities of such specialists and how much do they earn

Open Source is actively supported by both global technology giants (Google, Microsoft, IBM) and Russian players who contribute to the development of AI worldwide. According to the Strong AI in Industry Research Center (ITMO), in 2024, the top 10 Russian companies that create their own Open Source solutions or participate in other open source projects in the field of Data/ML include Yandex, Sber, T-Bank, Postgres Pro, VK, Avito, Evrone, MTS, Selectel and "Academy".

— A significant part of the events in Russian Open Source is related to the development of AI systems and language models. The main contributors here are Yandex and T-Bank. This bank, for example, recently made a large language model with 32 billion parameters publicly available and updated another one with 7 billion. And Yandex has discovered algorithms to speed up learning and compress language models, as well as a neural network—based development platform," says Evgeny Perov, product director at Compass corporate messenger. — The big news was VK's entry into the "big" Open Source. The company plans to open the source code of its products, IT systems, libraries for developers, and so on. This is positive news for the community — the market will be able to use the best practices. This will raise the overall quality level of IT products, both Open Source and proprietary.

According to Dmitry Ovchinnikov, head of the Gazinformservice Strategic Product Development laboratory, open source software can be used as a tool for testing penetration into information systems, improving the convenience of information security administration, or for highly specialized tasks. AI tools also allow you to automate the work of the support service and increase employee efficiency, for example, by creating assistants for developers when writing code," he notes.

— All over the world, open source promotes innovation, unique and powerful solutions. The progress of the Russian IT industry is linked to well-structured development processes, including Open Source, as well as the opportunity not to be disconnected from the international context," emphasizes Ruslan Gainanov, CEO of Team Force.

Open Source is shifting into AI

Today, the artificial intelligence market can be considered one of the drivers of the development of Open Source. This is facilitated by projects such as Hugging Face, PyTorch, and TensorFlow, which allow companies to implement AI faster without starting development from scratch. And the appearance of open models (for example, Llama 2 from Meta*, Qwen from Alibaba, Gemma from Google, YandexGPT 5 Lite, Ernie from Baidu) only strengthens the trend towards the availability of AI.

This allows startups and non-profit institutions operating on tight budgets to use cutting-edge technologies in their fields. Open Source also helps large players who do not specialize in AI to implement artificial intelligence into their products or business processes faster.

— As an AI researcher and an Open Source enthusiast, I am very interested in developing AI tools that would help in creating open source projects. So, at ITMO, we started creating an open OSA (Open Source Advisor) tool, which is aimed at helping research teams to output their research results in the form of reusable repositories," says Nikolai Nikitin, head of the ITMO frontier laboratory.

— We are seeing an increase in the number of startups and expanding access to high technologies in niches that previously either could not invest in AI at all or were generally limited in investments. These are, for example, social projects or individual industries, small companies, and so on," says Nikolai Nikitin. Or companies that have no experience working with AI, but are ready to use new technologies in their solutions to improve the quality of support work or increase employee efficiency," he points out.

— Open Source is becoming one of the key ways to overcome barriers in the development of AI. This approach significantly saves resources, speeds up product launches, and makes AI adoption more predictable. The Open Source community allows you to test ideas faster, share improvements, and bridge technological gaps through collective retraining and reuse of developments," says Sergey Berezhnoy, Director of Developer Relations at Yandex. — Open Source is a culture of collaboration, speed and efficiency. And this is exactly what helps AI become an affordable technology in the real economy.

How does Open Source affect the cost of development

Further training of models makes it possible to adapt the selected model to specific tasks and increase its effectiveness in a specific area. Despite the fact that basic LLMs that are already trained on large amounts of data can show good results, they are not always ideally suited for specialized tasks (for example, in medicine or finance), market participants note. The retraining process allows the model to gain additional knowledge based on specially selected data, after which it begins to better meet the company's objectives.

For example, Avito has trained the Mistral 7B neural network in Russian, adapting it to work with the company's ads. In 2024, T-Bank made two LLM models available at once — T-Pro and T-Lite. Both models have been further trained and adapted to the Russian language. According to the company, the Qwen-2.5 training is provided by Alibaba Group (based on them, T-Bank models have been created. — Izvestia) It allowed us to optimize development costs by 80-90% compared to learning from scratch. And in 2025, the open T-Pro 2.0 model with a hybrid reasoning mode was introduced. The total development costs, including the cost of computing power for R & D and final retraining, as well as salaries of employees, do not exceed 120 million rubles, the bank notes.

— We needed to find the right balance between fully pre-training our LLM models from scratch and using the latest Open Source models. Learning from scratch allows for complete customization, but it is redundant, difficult, and expensive. Open models may not meet the desired properties, but they are constantly improving and reducing the gap from their proprietary counterparts," says Anatoly Potapov, head of fundamental model development at T—Bank.

For high—quality retraining, the company's own data is used. These can be knowledge bases, existing regulations, orders and job descriptions, contracts and technical specifications.

How to save money using Open Source solutions

Thanks to the further education of Open Source, not only companies benefit, but also entire states. Developing countries also get a chance to create their own AI-based products and benefit economically from them. The way technology leaders like the United States and China do it. Among Chinese products, for example, the DeepSeek R1 neural network, released in early 2025, is fully open for commercial and research use.

DeepSeek has benefited from open research and Open Source (for example, PyTorch and Llama from Meta). They came up with new ideas and built them based on other people's work. Since their work is published and open, everyone can benefit from it. This is the power of open research and open source code," says Jan Lecun, Chief Artificial Intelligence Specialist at Meta*.

When Alibaba Cloud released more than 100 new open AI models, Jinren Zhou, the company's CTO, stated that "this initiative is designed to empower developers and corporations of all sizes to make better use of AI technologies and further drive the growth of the Open Source community."

Sergey Ponomarenko, Director of LLM products at MTS MWS, noted that the development of open LLM models in Russia will allow companies, both novice developers and researchers, to create solutions based on neural networks without investing significant resources in development and equipment.

At the end of last year, MTS MWS announced its intention to release an LLM-B2B Cotype Nano model with the ability to customize for specific tasks. The model itself was created based on Alibaba Cloud's Qwen 2.5 and upgraded by MTS MWS on various datasets, including synthetic ones, with simulations of real-world scenarios.

How to license Open Source

The further distribution and use of Open Source largely depends on the development of licensing. So, licenses appeared designed to protect developers from unfair use of their work. For example, if a large company makes changes to Open Source and uses it for commercial services, the license may require publishing these improvements in the public domain. Thus, the principle of "openness" of the project is preserved.

— Preference is given to the source code due to the lack of royalties for software licenses, which makes it possible to reduce the overall IT budget by gradually replacing the company's least critical services. Open Source solutions do not require business process restructuring, but, on the contrary, allow you to refine the software for existing ones," says Sergey Kharitonov, Technical Director of MTS Bank's IT cluster.

However, there are exceptions where, seeing the interest of the professional community, companies simplify the license terms, opening up more opportunities for them to research and experiment. For example, at the beginning of the year, Yandex made the pretrain version of YandexGPT 5 Lite publicly available, limiting its commercial use at the license level. Despite the fact that it was downloaded more than 15,000 times in less than a month, creating more than ten quantized models based on it and upgrading the instruct versions, the community was unhappy with the restrictions. Later, the company released the instruct version and updated the license. Now you can use the model for any purpose, including commercial ones, if the volume of output tokens does not exceed 10 million per month. For example, this number of tokens is enough to create and support chatbots on small and medium-sized sites, to generate product descriptions in online stores with a limited assortment, to automate responses to customers in service centers, or to analyze user reviews on sites with moderate traffic, the company notes.

Building a new AI world: Russia begins another digital breakthrough

But the legislative framework and human resources are not keeping pace with the dynamics of innovation.

Open Source AI continues to be actively used by both small businesses and large companies. Combined with the constant growth in the number of participants and the development of licenses, this creates the right environment for creating innovations and accelerating technological progress in general. Open technologies are increasingly becoming part of hybrid solutions, where the basic technology is free and accessible to everyone, while add—ons such as integration, support, and customization are paid for. This approach helps not only to build the reputation of a technology leader, solve the tasks of an HR brand and contribute to the development of the industry, but also brings direct benefits to the business.

*Meta company is recognized as an extremist organization in the Russian Federation