Startup Mostik Unveils Weight-Bridging Technique to Let AI Models Communicate Without Text

A newly emerged artificial intelligence startup called Mostik has introduced a mathematical technique that enables AI models to communicate directly through their underlying weight values rather than generating text prompts, according to reporting first published by Wired. Named after the Russian word for "bridge," the company was formed by a team of mathematicians seeking to transfer the capability of massive foundational models into smaller, more resource-efficient neural networks.

To demonstrate the approach, Mostik established a direct link between two open-weight models: the 753-billion-parameter GLM-5.2 model and a 4-billion-parameter edition of Qwen-3.5 designed to run on mobile devices. The resulting hybrid system operated at one-twentieth the running costs of the full GLM-5.2 setup while delivering benchmark performance positioned precisely halfway between the two original models. Mostik also applied the architecture to create a system that reached the top spot on ARC-AGI 3, a widely recognized artificial intelligence evaluation contest.

Mostik Chief Executive Officer Sasha Malysheva, who created the communication framework, noted that combining outputs from multiple models routinely yields stronger capabilities than relying on single architectures. Drawing an analogy to a well-known mathematical concept where averaging multiple crowd estimates produces a more accurate guess of a pig's weight than an individual expert, Malysheva told Wired that the sector's future lies outside monolithic models and compute scaling, focusing instead on interlinked networks.

If adopted broadly, direct model-to-model communication could significantly bolster the efficiency of open-weight models competing against closed, proprietary systems built by frontier research organizations such as OpenAI and Anthropic. Vladimir Arustamian, tech lead at AI software firm Lovable, remarked that Mostik's technology could allow specialized domain models in subjects like biology and physics to pair seamlessly with frontier systems. Arustamian added that the startup deployed a functional prototype in a matter of months, achieving progress he previously expected would take years.

Former Google DeepMind computer scientist Karl Tuyls noted that Mostik's framework allows developers to approach large-model outputs without requiring the full model to handle the complete operational loop, describing the method as an obvious fit for teams optimizing execution efficiency. Meanwhile, Mostik Chief Scientist Stanislav Smirnov—a professor at the University of Geneva and a 2010 Fields Medalist—explained that establishing a formal mathematical language between separate AI models remains an open challenge, making Mostik’s weight-bridging method an effective workaround.

Smirnov further suggested that deeper mathematical analysis into how models communicate could provide new insights into AI reasoning processes and how they compare to human cognition. Malysheva, who entered competitive mathematics at a top institution in St. Petersburg after being challenged by her older brother, noted she advanced the bridging architecture despite initial skepticism from industry peers who warned the mathematical hurdle would be too steep.