Open-Source AI Models: Finding the Balance Between Openness and Sovereignty

Arutyun Avetisyan, an academician and director of the Ivannikov Institute for System Programming of the Russian Academy of Sciences, moderated the session. He outlined its main goal: to jointly determine the way forward in an era when AI models are becoming ubiquitous.

Opening the discussion, Francesco Stabilito, a digital technology and international development specialist at the United Nations Office for Digital and Emerging Technologies (Italy), emphasized that open-source software is becoming a key tool for technological sovereignty. However, access to models alone is not enough.

“The beauty of open source is that it lowers the barriers to entry. But we know that access alone is not enough. It is better to start with the problem you need to solve rather than think about the technology.”

Dmitrii Markov, CEO of VisionLabs (Russia), presented the perspective of Russian businesses. He said Russia has adopted a risk-based approach: controlled, sovereign models are used for critical infrastructure, while open-source solutions are acceptable in the commercial sector.

“If you are using a product for free, you are the product. Since you cannot control this process, many countries are developing their own models for critical infrastructure,” Markov noted.

Andrey Neznamov, managing director of Sber Center for Human-Centred AI and a member of the UN Independent International Scientific Panel on AI, focused on the need to rethink the concept of AI sovereignty. In his view, it encompasses an entire ecosystem: deploying technology in industry, training specialists, and securing access to data and computing resources. As an example of practical efforts, Neznamov described Russia’s contribution to the global AI agenda. Sberbank and Yandex have made their flagship models, GigaChat, Kandinsky, and Alice (Alisa), available to users, helping bridge the digital divide, particularly in BRICS countries. International cooperation is also expanding: an AI alliance bringing together 26 countries has been established. Russia also has a National Code of Ethics for AI, with organizations from 60 countries participating.

Monis Ali Zain Ul Abid Edin, an AI server-side engineer at Facehawk Limited (Pakistan), drew on his experience developing AI for public safety to emphasize the importance of risk assessment. He said fine-tuning is preferable when a suitable base model is available. However, developing a model from scratch is necessary when strict licensing restrictions and privacy requirements apply, particularly in the public sector.

Sharing practical experience with AI deployment in Egypt, Mostafa Ahmed El Demery, an analytics and machine learning engineer at Telecom Egypt, said training data must be adapted to the country’s specific cultural and linguistic context.

“Train your model on local-language data. In Egypt, we have many dialects, and if you are going to train a model on a specific dialect, you need to specify which one you intend to use, because the model will not know your particular dialect,” the engineer noted. He also shared an idea for creating an open-source AI model that would gather information from across the Middle East.

Trevor Takudzwa Mudondo, product manager at Econet AI Zimbabwe, stressed that applications built on open-source technology are particularly important for Africa when they address local challenges in healthcare, education, and agriculture. Developing foundational models from scratch is too expensive, making open source the best option.

Mohamed Ben Amor, director general of the Arab Information and Communication Technologies Organization (AICTO), based in Tunisia, discussed the role of open-source models in implementing regional strategies. He said Arab countries have moved toward more deliberate AI adoption by developing national strategies. AICTO has developed an Arab AI strategy, an ethics pact, and an action plan. Open source, he argued, enables countries in the region to safeguard their technological sovereignty and protect personal data.

International participants showed considerable interest in Russian models. In response to a question from Mostafa Ahmed El Demery, Andrey Neznamov confirmed: “Russian open-source models are competitive. Most importantly, we make the model weights available, allowing any country to localize a model without data leaking abroad. This is important for AI sovereignty.”

Trevor Takudzwa Mudondo was impressed by the prospect: “We have young developers working in AI. I hope we will add your models, Alice and GigaChat, to our toolkit.”

The experts agreed that only a handful of countries and organizations have the resources to develop foundational technologies independently; the rest of the world should seek to cooperate. As Arutyun Avetisyan noted, foundational technologies should be equally accessible to everyone. Russia has proposed creating an international repository of trusted AI solutions, an initiative that has already received support from partners in India.

“We need to move from platforms for discussion to platforms where our engineers work together to create value and develop new technologies,” Avetisyan concluded.

The participants agreed to continue the dialogue and work together to develop open, secure AI solutions that can be adapted to local conditions.