Sber's GigaChat: How Russia's Largest Bank is Building a Cutting-Edge AI Language Model from Scratch

Sber, Russia's dominant bank and tech powerhouse, is developing GigaChat, a groundbreaking neural network language model entirely from scratch.

Anton Frolov, Head of Sber's Generative AI Development Division, shared insights in a recent Q&A on the project's inception, challenges, and vision.

Why Train from Scratch?

By focusing on in-house pre-training, Sber ensures complete control over GigaChat's foundational knowledge, avoiding risks associated with adapting foreign open-source models.

This resource-heavy approach prevents potential disruptions like access cutoffs and preserves the model's core capabilities intact.

Advanced Features Boosting User Engagement

The updated GigaChat features long-term memory that recalls users' interests, professions, and habits across sessions.

Its voice mode enables seamless real-time conversations, complete with automatic internet searches, interruptions, and unwavering context retention.

Overcoming Trust Barriers in Russia

In Russia, 25% of the working-age population uses AI assistants, though skepticism persists, largely due to prejudice rather than poor experiences.

Trust surges among users who have been helped, with satisfaction rates jumping 26 to 46 percentage points.

Global Lessons and Future Horizons

Inspired by efficient Chinese open-source models like DeepSeek and Qwen, Sber commits to releasing its models openly to build a robust ecosystem.

Over the next three to five years, GigaChat will evolve into a universal assistant, replacing dozens of apps and guiding decisions on health, finances, and careers.

Prospects for Russian-Chinese AI collaboration include joint R&D and knowledge sharing to propel global progress.