Russian Startup Teaches AI Models to Understand Each Other Without Words
Startup Mostik has developed a way to combine the capabilities of different AI models without exchanging text responses. Instead of lengthy back-and-forth exchanges, the neural networks share mathematical values from their internal representations.
Models are typically connected sequentially: one generates an answer, and another takes it as input, which requires time and computing resources. Mostik’s method transfers information between models at a deeper level, allowing a smaller neural network to tap into the capabilities of a much larger one. To demonstrate the approach, the team connected the Chinese GLM-5.2 model, with 753 billion parameters, to the compact Qwen-3.5 with 4 billion parameters, which can run on a mobile device. According to the developers, the resulting system performed exactly halfway between the two original models while costing 20 times less to run than the full GLM.
Mostik also put another of its systems at the top of the ARC-AGI 3 competition, which tests AI’s ability to solve novel problems. The team has not yet disclosed details.
Sasha Malysheva, the startup’s CEO, believes the future of AI may belong not to a single giant model, but to a collective of specialized neural networks that could be combined with a large model, for example, to handle medicine, physics or biology.
