Russian startup Mostik creates bridge allowing AI models to share knowledge without text

Russian mathematicians at the startup Mostik have introduced a technique that enables artificial-intelligence models to “talk” by sharing the numerical values of their weights rather than generating text. The method allows a powerful, large-scale model to transfer part of its ability to a much smaller model, dramatically reducing computational expense. In a demonstration, Mostik combined the 753-billion-parameter GLM-5.2 model with a 4-billion-parameter Qwen-3.5 model, creating a hybrid that costs one-twentieth as much as the full GLM and performs at roughly the midpoint between the two.

CEO Sasha Malysheva likens the concept to ensemble learning and argues it could help open-weight models compete with closed systems from firms like Anthropic and OpenAI. Tech lead Vladimir Arustamian and former DeepMind researcher Karl Tuyls praised the speed of development, while chief scientist Stanislav Smirnov noted the mathematical challenges of finding a common language between models. The team hopes the breakthrough will spur more specialized, domain-specific AI without relying on ever-larger monolithic architectures.