Russian mathematicians at startup Mostik have developed a technique that allows different artificial intelligence models to communicate without exchanging text. Instead, the approach connects models through mathematical values associated with their weights, potentially allowing smaller AI systems to benefit from the capabilities of much larger ones, tells Wired.com (full article available to subscribers).
Mostik, whose name means bridge in Russian, was cofounded by CEO Sasha Malysheva, who developed the approach. Rather than passing the text output of one model to another, a process that can consume considerable time and computing resources, Mostik creates a mathematical bridge between models.
The company has already used the technique to develop a system that reached the top of ARC-AGI 3, a challenging benchmark designed to test AI reasoning. Mostik has not disclosed details of that system because the competition is ongoing.
For another demonstration, the researchers connected two Chinese open-weight models: GLM-5.2, with 753 billion parameters, and a 4-billion-parameter version of Qwen-3.5 capable of running on a mobile device. The resulting hybrid delivered performance halfway between the two models while costing just one-twentieth as much as running the full GLM model.
The approach builds on the established idea that groups, or ensembles, of AI models can outperform individual systems. Mostik aims to make such collaboration considerably more efficient. The technology could eventually allow powerful general-purpose models to work alongside smaller models trained for specialized fields such as biology or physics.
Malysheva believes AI development may shift away from relying primarily on increasingly large, monolithic models toward combinations of specialized systems.
Mostik’s chief scientist, University of Geneva professor and Fields Medalist Stanislav Smirnov, sees another possibility. Studying the mathematical connections between models could reveal more about the mechanisms behind AI reasoning and perhaps identify similarities between artificial intelligence and the way humans solve difficult problems.
