Russian math tutor’s artificial intelligence model communicates without language – knowledge from large models can be effectively transferred to small models

Russian mathematicians from startup Mostik have developed a method that allows artificial intelligence models to exchange information without text messages – using mathematical values ​​embedded in their scales. In practical terms, this means that the functionality of a larger model can be transferred to a smaller model to more effectively increase its intelligence.

Image source: ChatGPT

The startup’s name reflects how the technology works. Rather than sequentially transmitting textual responses from one model to another, as is currently the case, Mostik’s proposed system uses mathematical “bridges” to connect their internal numerical representations. The “bridge” must convert one model’s internal numerical representation into a format that the other model can use – no text appears between the models at all, and neither LLM is retrained.

Mostik said it has used this approach in the ARC-AGI-3 system, a competition that tests the ability of artificial intelligence to generalize rules and solve new abstract problems. According to the company, its solution has become a leader in current ratings.

To demonstrate the method, Mostik combined two open-scale Chinese models: GLM-5.2 with 753 billion parameters and a version of Qwen-3.5 with 4 billion parameters that runs on mobile devices. The resulting hybrid system costs about 20 times less than the flagship GLM, placing the result squarely between the two original models.

Sasha Malysheva, the head of the new startup that came up with the idea, believes that combining multiple models may be more promising for the development of artificial intelligence than continuously expanding the size of the model. Meanwhile, Vladimir Arustamyan, technical director at artificial intelligence company Lovable, noted that the ability to connect general models with specialized systems in biology, physics and other fields could spur the creation of more such models.

“Mostik technology means you can approach the quality of a large model without requiring the large model to handle the entire cycle, which allows for significantly improved results using only small models running in parallel.”” said Carl Tuyles, a former computer scientist at Google DeepMind who is familiar with the company’s technology. According to Tuyles, this approach is the obvious choice for anyone looking to run a model as efficiently as possible.

Mostik’s chief scientist is Stanislav Smirnov, a professor at the University of Geneva and the 2010 Fields Medal winner. Finding a common mathematical language for the two AI models has been unexpectedly difficult, he said, but further research into the method could help better understand the principles of AI and its possible similarities to human thinking. If this technology becomes widely available, it could add value to open weight models, allowing them to better compete with closed proprietary models from leading labs like Anthropic and OpenAI.