The startup Mostik has developed a method for AI models to exchange internal states directly, bypassing the need for intermediate text generation. The system utilizes an adaptable module, or bridge, that converts the internal representation of one model into a format compatible with another, without modifying the original model parameters.
During tests, the team connected the 753 billion-parameter GLM-5.2 model by Z.ai with the 4 billion-parameter Qwen-3.5 model by Alibaba. In this setup, the larger model processed the query, while the smaller model generated the final text. This approach reduced computational requirements by approximately 2.5 times compared to a single medium-sized model.
According to Mostik, this method closed about half the quality gap between small and large models, providing a 10 percentage point improvement over standard text-based interaction at the same computational cost. The team also applied this technique to the ARC-AGI-3 benchmark, achieving high performance, though these results remain preliminary.
Mostik researcher Stanislav Smirnov noted that internal representations across different neural networks are not directly comparable, as they encode information differently. While the current results are based on internal experiments, experts suggest this approach could enable the combination of universal models with specialized systems for specific fields like biology or physics.