The architecture of modern large language models (LLMs) assumes that when working together, neural networks communicate exclusively through text generation. This is a colossal bottleneck: during query processing, the model generates over a hundred hidden vectors—about a million numerical values, equivalent to roughly 2 MB of internal state. Only a tiny fraction of this information makes it into the final text response. The team at startup Mostik proposed an elegant solution to this problem by creating a separate adaptable "bridge" module that translates one model's internal representations directly into another's computational space, bypassing the verbalization stage.
A key feature of the approach is the complete static nature of the interacting neural networks' parameters. Only the intermediate module, which learns to map the latent spaces of different architectures, is subject to adjustment. This opens up opportunities for creating hybrid systems without expensive fine-tuning.
Experiment: An alliance of giant and tiny
To verify the method, researchers connected the powerful GLM-5.2 from Z.ai (753 billion parameters) with the modest Qwen-3.5 from Alibaba (4 billion parameters). In this pairing, the "heavy" model handled exclusively query analysis without generating a response. Its internal state was passed through the "bridge" to the small model, which then produced the final text.
The results are impressive: this configuration reduced the quality gap between the small and large models' responses by about half. At comparable computational costs, the hybrid system showed a 10-percentage-point advantage over the classic scheme with text-based data transfer. Moreover, the GLM-5.2 and Qwen-3.5 pairing proved roughly 2.5 times more computationally efficient than a monolithic mid-sized model of similar quality.
In search of a common mathematical language
Mostik's Chief Scientific Officer Stanislav Smirnov emphasizes the fundamental complexity of the task: the internal representations of different neural networks are not directly comparable. Even when solving the same problem, models encode information in their hidden states in fundamentally different ways. "It seems that a suitable mathematical language does not yet exist," he notes. Studying such correspondences, in his view, could shed light on how AI systems represent knowledge and make decisions, as well as reveal common patterns in their "reasoning."
The practical value of the development is obvious for the industry. Former Google DeepMind researcher Karl Tuyls sees a scenario where a resource-intensive model is used only for "thinking," while a compact model handles text generation. Another promising direction is combining a general-purpose model with specialized experts in biology, physics, or finance. Lovable's Technical Lead Vladimir Arustamyan believes this will increase demand for narrowly specialized models, allowing complex systems to be assembled from ready-made components.
Baptism by fire: ARC-AGI-3
Mostik also applied its method to the ARC-AGI-3 test—a set of tasks requiring AI to 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 due to the ongoing competition.
However, these claims should be treated with caution: the data in ARC-AGI-3 has "preview" status and is based on incomplete tests. All other performance metrics—reducing the gap between models and lowering costs—are based on the team's internal experiments and have not yet received independent confirmation.
My analysis: This work marks an important shift from the "communication through text" paradigm to direct exchange of hidden states. If Mostik's results are confirmed by independent tests, we stand on the threshold of truly modular AI systems where "brain" and "speech" can be separated and optimized individually. This is not just optimization—it is a new perspective on artificial intelligence architecture, where knowledge is transmitted at the level of mathematical abstractions rather than words.