A Russian startup called Mostik has developed a technique that lets artificial intelligence models communicate by sharing the mathematical values embedded in their weights, rather than exchanging text output. According to a recent Wired profile, this approach allows a larger model's capabilities to be transferred to a smaller one far more efficiently than existing methods. The company's name translates to "bridge" in Russian, reflecting how the technology connects different models.

To showcase the concept, Mostik built a bridge between two Chinese open-weight models: a 753-billion-parameter version of GLM-5.2 and a 4-billion-parameter version of Qwen-3.5 small enough to operate on a mobile device. The hybrid system that resulted costs just one-twentieth of running the full GLM model, while delivering performance that sits exactly midway between the two original models. The startup also applied its method to reach the top of ARC-AGI 3, described as a notoriously tough competition for AI models, though the team declined to share details because the contest is still active.

CEO Sasha Malysheva, who created the approach, told Wired that merging multiple model outputs typically produces better results than relying on a single model—a principle she compared to the math puzzle where random people's averaged guesses about a pig's weight beat an expert's estimate. Traditional ensemble methods feed one model's output into another, which consumes significant time and expense, she explained. Mostik's chief scientist, Stanislav Smirnov—a University of Geneva professor and 2010 Fields Medalist—noted that identifying common ground between two AI models is surprisingly hard, saying "there seems to be no appropriate mathematical language yet."

The technique could reshape how AI advances by making open-weight models more competitive with the proprietary systems from frontier labs like Anthropic and OpenAI. Malysheva told Wired she doesn't believe future progress will come from monolithic models or pure scaling—the strategy of building ever-larger models fed with more data. Instead, she suggested, combining many different models may prove a superior path forward. If Mostik's bridges can pair frontier models with specialized ones focused on domains like biology or physics, it would incentivize training far more niche models, said Vladimir Arustamian, tech lead at AI software company Lovable, who called the achievement something he would have guessed was years away. Karl Tuyls, a former Google DeepMind computer scientist familiar with the technology, said it's a no-brainer for anyone running models efficiently, since you can approach large-model quality without the large model handling the entire loop, gaining substantial improvements with just a smaller model alongside.

The approach represents a fundamental shift in how AI systems might collaborate, moving from sequential text exchanges to direct mathematical communication. By allowing smaller models to tap into larger ones' strengths without the cost of running those behemoths in full, Mostik's bridges could democratize access to high-performance AI while slashing computational expenses. For companies deploying AI at scale, the ability to cut costs by a factor of twenty while maintaining halfway performance marks a compelling trade-off between efficiency and capability. Enterprises betting on domain-specific intelligence rather than general-purpose giants may find this architecture opens doors that pure scaling never could, particularly where budget constraints or mobile deployment demands lighter footprints.