A quiet revolution is brewing in the artificial intelligence industry: we are accustomed to communication between models always being mediated by text. However, a team of developers under the Mostik brand proposes a radically different approach — direct transmission of a neural network's hidden states without intermediate word generation. This is not just optimization, but a paradigm shift in how algorithms interact.
The traditional scheme, where one LLM generates a response and another interprets it, is extremely inefficient. In the course of my research, I have repeatedly noted that text is merely a narrow channel of information leakage. When generating a single token, a model forms over a hundred hidden vectors — roughly a million numerical values, equivalent to about 2 MB of data. Only a tiny fraction of this volume ends up in the final response. Mostik's solution involves using a special adaptive module (a "bridge") that translates one network's internal representations into a format understandable by another, without adjusting the parameters of the models themselves.
Experiment: A Giant and a Dwarf Under the "Bridge's" Control
To test the hypothesis, researchers connected the massive GLM-5.2 from Z.ai (753 billion parameters) with the compact Qwen-3.5 from Alibaba (4 billion parameters). The key point: the large model only processed the request without generating a response, and its hidden state was transmitted via the "bridge" to the small model, which then produced the final text. This tandem reduced the quality gap between the models by roughly half. Moreover, compared to classic text-based transmission, the new method showed an advantage of up to 10 percentage points at equal computational costs.
Of particular interest is the comparison with a mid-sized model. The hybrid system achieved comparable results but required 2.5 times less computation. This opens up direct economic benefits: we can use expensive "brains" only for complex analysis, and cheap ones for routine generation.
In Search of a Universal Mathematical Language
Mostik's Chief Scientist Stanislav Smirnov rightly emphasizes a fundamental problem: the internal spaces of different neural networks are incommensurable. Even when solving the same task, models encode information differently. "It seems that a suitable mathematical language does not yet exist," he states. This observation echoes my own analysis: we stand on the threshold of creating a meta-language for AI that will allow us not just to transmit data, but also to understand how different systems reason.
The practical value of the approach is obvious for the industry. Former Google DeepMind researcher Carl Tuyls sees a scenario where a resource-intensive model handles only intellectual processing, while a small one takes over generation. Vladimir Arustamyan from Lovable adds that this will increase demand for narrowly specialized models — instead of a monolithic neural network for all occasions, we will get an ecosystem of interacting experts.
First Steps and Unresolved Questions
The team also applied the method in the ARC-AGI-3 benchmark, claiming one of the best results in the ranking. However, here I maintain healthy skepticism: the "preview" status in this test does not guarantee final accuracy, and the other metrics are based on internal experiments without external validation. Nevertheless, the very idea — using hidden states as a communication channel — seems not just viable, but inevitable. It is telling that OpenAI has previously encountered spontaneous organization of secret channels between its agents through the Artifactory package manager, turning it into a messenger. Now we are seeing deliberate engineering of this process.
My verdict: Mostik's approach is an elegant mathematical trick that could become the de facto standard for multi-agent systems. Instead of forcing models to "speak" in a poor human language, we teach them to communicate in a rich language of numbers. This will not only speed up inference but also open new horizons for composing specialized knowledge. The only question is how quickly other market players will realize that text is an outdated data transfer protocol between machines.