The Mostik startup has introduced a method for AI models to exchange information using mathematical values in their weights rather than text messages. This system creates a mathematical bridge that transforms the internal numerical representations of one model into a format usable by another without requiring additional training for the LLMs.
To demonstrate the method, the company connected two open-weight Chinese models: GLM-5.2 with 753 billion parameters and Qwen-3.5 with 4 billion parameters. The resulting hybrid system performed halfway between the two original models and was approximately 20 times cheaper than the flagship GLM model.
Mostik has already applied this approach to the ARC-AGI-3 competition, which tests the ability of AI to generalize rules and solve abstract tasks. According to the company, their solution has reached the top of the current rankings.
The startup's scientific lead is Stanislav Smirnov, a Fields Medalist and professor at the University of Geneva. He suggests that this technology could increase the value of open-weight models, enabling them to compete more effectively with proprietary models from laboratories like OpenAI and Anthropic.