Russian Mathematicians Teach Neural Networks to Talk to Each Other
Russian startup Mostik has proposed a new way for AI models to work together that could reduce the need to continually scale up computing power.
There are a huge number of neural networks in the world, each designed to handle its own set of tasks. But could large language models interact with one another to produce more comprehensive answers? A team of Russian IT professionals has taken on that challenge.
Long Exchanges Are No Longer Necessary
The team at Russian startup Mostik has found a way for AI models to interact without directly exchanging text messages. Instead of lengthy back-and-forth exchanges, the neural networks “communicate” through mathematical values in their internal representations. This approach has a significant advantage over conventional methods of interaction. Typically, models work sequentially: one generates an answer, and another receives it. That process takes time and requires substantial computing resources.
The method developed by Mostik allows information to be transferred between models at a deeper level, helping a smaller neural network tap into the capabilities of a much larger one.
Early Practical Results
To demonstrate that the idea works, the Russian startup’s team built a bridge between two Chinese open-weight models: the largest version of GLM-5.2, with 753 billion parameters, and a 4-billion-parameter version of Qwen-3.5 that can run on mobile devices. The resulting hybrid was 20 times cheaper than the larger model, while the quality of its responses fell somewhere between the performance levels of the two original neural networks. The startup used the same approach to build a model that quickly moved to the top of the exceptionally difficult ARC-AGI 3 AI model competition.
“In machine learning, it is well known that combinations of models produce better results than individual models,” said Alexandra Malysheva, CEO of Mostik.
Don’t Expect an “Electronic Superintelligence”
The concept proposed by the Russian startup’s team could significantly increase the value of open models and help them compete more effectively with closed AI platforms offered by leading labs such as Anthropic and OpenAI.
Malysheva believes that the future of the AI industry lies precisely in interaction among different models. She doubts that an “electronic superintelligence” will ever emerge that can handle absolutely every task.
Stanislav Smirnov, the startup’s chief research scientist, explained that finding common ground between two AI models is exceptionally difficult. There is simply no suitable mathematical language for doing so. In the future, however, Mostik’s work could go a long way toward closing that gap.
Balancing Quality and Cost
Notably, the startup’s research reflects a broader trend in neural-network development in recent years. Developers are looking for ways to improve model performance without endlessly scaling up computing power. In 2024, Sber published the weights of a smaller GigaChat model based on a Mixture of Experts architecture. Although it has 20 billion total parameters, the platform uses about 3.3 billion active parameters in a single pass. It is a clear example of how a company sought to reduce computing costs by using only the model resources needed at a given moment.
That same year, T-Bank opened up Russian-language models with 7 billion and 32 billion parameters, built on Qwen 2.5 and additionally trained in Russian. This was another attempt to find a balance between the performance of AI infrastructure and its cost.
Leading global technology giants are also regularly looking for ways to transfer the capabilities of large language models to more compact platforms.
Of course, Mostik’s work is still at a very early stage, and its specialists have many technical questions to resolve. But if the approach scales and the reported performance figures hold up, it could lower the computing cost of generative AI and make architectures built from multiple specialized models a competitive alternative to a single enormous LLM. This is particularly important for Russia, which is working to build a competitive AI architecture with limited computing resources.
