A quiet revolution is brewing in the artificial intelligence industry: we are accustomed to communication between models always being mediated by text. However, a development team under the Mostik brand proposes a radically different approach—direct transmission of a neural network's "thoughts," bypassing the verbalization stage. Instead of generating intermediate text, a special adaptive module is used, which I call a "bridge." It converts the internal vector representations of one model into a format understandable to another, while the parameters of the systems themselves remain untouched.
This is a fundamental shift. When generating even a single token, a large language model produces over a hundred hidden vectors—about a million numerical values, equivalent to roughly 2 MB of data. Only a tiny fraction of this information makes it into the textual response. The new channel aims to transmit precisely these rich intermediate states.
Experiment: A giant thinks, a dwarf speaks
To test the hypothesis, researchers assembled a pairing of GLM-5.2 from Z.ai (753 billion parameters) and Qwen-3.5 from Alibaba (4 billion parameters). The key point: the large model processed the query but did not generate a response. Its internal state was transmitted via the "bridge" to the small model, which then produced the final text.
The results are impressive. This pairing closed about half of the quality gap between the small and giant models. Compared to classical textual transmission, the new method showed an advantage of up to 10 percentage points at identical computational costs. Moreover, a hybrid system comparable in quality to a mid-sized model required approximately 2.5 times less computation.
In search of a common mathematical language
Mostik's chief scientific officer, Stanislav Smirnov, rightly notes a fundamental problem: the internal representations of different neural networks are not directly comparable. Even when solving the same task, models encode information in their hidden states differently. "It seems that a suitable mathematical language does not yet exist," he states. This opens up a new area of research: studying correspondences between latent spaces could shed light on how AI represents knowledge and makes decisions.
The practical value of the approach is obvious. Former Google DeepMind researcher Karl Tuyls sees a scenario where a resource-intensive model is used only for "thinking," while text generation is delegated to a compact system. Another path is the symbiosis of a general-purpose model with specialized ones (biology, physics), which eliminates the need to train monolithic systems on all data at once.
A bid for leadership in ARC-AGI-3
The developers applied the method to the challenging ARC-AGI-3 benchmark, 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—the competition is ongoing. However, these data should be treated with caution: the "preview" status means an incomplete set of tests, and other metrics are based on the team's internal experiments.
My analysis: This is an elegant solution to a long-overdue problem. The AI economy is hitting the cost of inference, and the ability to "divide labor" between models is not just optimization but a step toward modular intelligence architecture. However, the path from demonstration to industrial implementation is long: we still need to learn how to build universal "bridges" between any architectures, not just between specific pairs.