A quiet revolution is brewing in the artificial intelligence industry: we have grown accustomed to thinking that communication between models is always text generation. However, the development team behind the Mostik project has proposed a radically different approach, allowing neural networks to exchange "thoughts" directly, bypassing the stage of linguistic formulation.
The essence of the method lies in using a special adaptive module—a "bridge." This component translates the internal vector representation of one model into a format that another can interpret, without the need to convert information into words. Critically, the parameters of the neural networks themselves remain unchanged during such interaction—only the connecting element is adjusted.
Why does this matter? According to expert estimates, before generating even a single token, a large language model forms more than a hundred hidden vectors—this amounts to about a million numerical values or roughly 2 MB of data. Only a tiny fraction of this information makes it into the textual response; most of the "thought process" is simply lost. The new communication channel is designed to transmit precisely these rich intermediate representations.
The Experiment: Giant and Dwarf
To test the hypothesis, researchers linked the most powerful model, GLM-5.2 from Z.ai (753 billion parameters), with the compact Qwen-3.5 from Alibaba (4 billion parameters). In this pairing, the "giant" processed the request but did not generate a response, passing its state through the "bridge," while the final text was produced by the small model.
The results are impressive. This pairing reduced the quality gap in responses between the small and large models by roughly half. Moreover, compared to traditional text-based transmission, the new method showed an advantage of up to 10 percentage points at identical computational costs. A comparison with a mid-sized model also favored the hybrid system: it required 2.5 times less computation for a comparable result.
In Search of a Common Mathematical Language
Mostik's Chief Scientific Officer, Stanislav Smirnov, emphasizes a fundamental challenge: the internal representations of different architectures cannot be directly compared. Even when solving the same task, models encode information in their hidden states in completely different ways. "It seems that a suitable mathematical language does not yet exist," he notes. Studying these correspondences could not only improve system interaction but also shed light on how AI arrives at decisions in the first place.
The practical potential of the method is enormous. Former Google DeepMind researcher Carl Tuyls sees the main scenario in resource savings: an expensive model can be used solely for analysis, while generation is delegated to a small one. Another path is combining a general-purpose model with specialized ones (e.g., in biology or physics), avoiding the costly training of a single giant network on all data at once.
Trial by Fire and Skepticism
The team also applied the approach to the ARC-AGI-3 benchmark—a set of tasks where AI must adapt to new rules in real time. The system with the "bridge" showed one of the best results in the current ranking, although details are not yet disclosed, and the data has "preview" status and has not been externally verified. All other reported metrics are also based on internal experiments.
My analysis: This is an elegant solution that challenges the dominant paradigm of "text as a universal interface." However, it is worth remembering that so far we have observed OpenAI's AI agents creating hidden communication channels through package managers. If Mostik's "bridges" truly work, we stand on the threshold of a new era where AI efficiency will be determined not by size, but by the quality of inter-intelligence interaction. But independent, reproducible tests are needed for confirmation.