A quiet but significant shift is brewing in the artificial intelligence industry. We have grown accustomed to collaboration between neural networks occurring through the generation of intermediate text—a kind of "translation" of thoughts into human language for transmission to another algorithm. However, a team of developers under the Mostik brand proposes a radically different path: direct exchange of internal states, bypassing the verbal stage.
The essence of the method lies in using a special adaptable module, which I call a "bridge." It transforms the hidden vectors of one model into a format understandable to another, while the parameters of the neural networks themselves remain untouched. This is fundamentally important: we do not fine-tune the models, but merely create an intermediary layer for their seamless interaction.
Numbers and experiments: testing on giants
The developers claim that when generating a single token, a large language model forms over a hundred hidden vectors—this amounts to about a million numerical values or roughly 2 MB of data. Only a small fraction of this richness ends up in the textual response. Their transmission channel targets precisely these rich internal representations.
The empirical validation looks impressive. In a pairing of the giant GLM-5.2 from Z.ai (753 billion parameters) and the compact Qwen-3.5 from Alibaba (4 billion parameters), the large model processed the query but did not generate a response. Instead, its internal state was transmitted via the "bridge" to the small model, which then produced the final text. Such a tandem, according to the authors, closed about half of the quality gap between the models. When compared with classical text transmission, the new method showed an advantage of up to 10 percentage points at equal computational costs. Moreover, the hybrid system required approximately 2.5 times less computation than a medium-sized model with a comparable result.
The main problem: finding a common "language"
The key difficulty, acknowledged by the researchers themselves, is the lack of a unified mathematical space for different architectures. Mostik's Chief Scientific Officer, Stanislav Smirnov, rightly notes: "It seems that a suitable mathematical language does not yet exist." The internal encodings of information in different networks can differ radically even when solving identical tasks. This is precisely why the "bridge" must be adaptable—it learns to translate from one "dialect" of hidden states to another.
The practical value of the approach is obvious: we can use a heavy model only for complex cognitive work, while delegating response generation to a lightweight one. This opens the way to creating hybrid systems that combine universal and narrowly specialized models without the need to train a monster on all data at once.
Assessment and skepticism
Despite the intriguing results on the ARC-AGI-3 tests, where the system showed one of the best results, healthy skepticism should be maintained. All stated metrics are based on the team's internal experiments and have not yet received external confirmation. Nevertheless, if the method is confirmed, we stand on the threshold of a new paradigm in AI scaling, where efficiency is achieved not by increasing parameters, but by intelligent communication between specialized agents.
My verdict: this is a bold and potentially groundbreaking step, but it is far from production readiness. The key question is how stably the "bridge" will work across heterogeneous architectures and tasks outside laboratory conditions. We are watching the development closely.