Mostik has developed a new method for AI models to communicate without text, using a bridge module to transmit internal states directly.
The startup Mostik has launched a groundbreaking method that allows artificial intelligence models to share their internal states without generating intermediary text. This innovative approach replaces traditional text-based communication with a specialized adaptable module, known as a "bridge," which transforms the internal representations of one model into a format comprehensible by another.
Normally, when multiple neural networks collaborate, they rely on generated text as a means of communication. However, Mostik's method retains the original parameters of the models during interaction. According to the developers, before producing a single token, a large language model (LLM) generates over a hundred hidden vectors, equating to about a million numerical values, or roughly 2 MB of internal state. The new channel is specifically designed to transmit these intermediate representations efficiently.
Demonstrating the New Method
To validate their technique, the Mostik team connected the GLM-5.2 model from Z.ai, which has 753 billion parameters, with Alibaba's Qwen-3.5 model, containing 4 billion parameters. In this setup, the larger model processed requests without issuing a response, instead transmitting its internal state through the bridge, while Qwen-3.5 generated the final text output.
This connection reportedly reduced the quality gap between the smaller and larger models by approximately half. The new method also outperformed traditional text transmission by achieving a 10 percentage point improvement while requiring the same amount of computational resources. Furthermore, when compared to a medium-sized model, the hybrid system of GLM-5.2 and Qwen-3.5 necessitated about 2.5 times fewer computations to achieve similar results.
Exploring Model Communication
Stanislav Smirnov, Mostik's chief scientist, emphasized the challenge of comparing internal representations across different neural networks. He noted that even when models address the same issue, their encodings can differ significantly. Smirnov suggested that a suitable mathematical language to bridge these differences is not yet available, but analyzing these correspondences may enhance our understanding of AI reasoning.
"It seems that an appropriate mathematical language does not yet exist," he stated in an interview.
Karl Tuyls, a former researcher at Google DeepMind, highlighted the practical applications of this approach, suggesting that resource-intensive models could focus solely on processing requests while smaller models handle response generation. This strategy could lead to the integration of universal models with specialized ones, enabling them to work together without the need for a single large model trained on diverse datasets.
Vladimir Arustamyan, technical director at Lovable, noted that this method could boost demand for specialized models as developers might prefer to combine different systems rather than creating a one-size-fits-all neural network.
Currently, these ideas are largely theoretical, as Mostik has only demonstrated the method using specific model pairs.
Application in Complex Tasks
Mostik has also applied its communication method in the ARC-AGI-3, which involves complex interactive tasks requiring AI to adapt to new rules independently. The team claims that their system achieved one of the best results in current rankings, although detailed results have not yet been disclosed as the competition is ongoing. It is important to note that these results are based on preliminary tests and have not been externally validated.
Earlier this year, OpenAI revealed a secret communication channel between its AI agents, which accessed the open internet, effectively creating an internal messaging system for its models.
