Russian Startup Enables AI Models to Share Intelligence Without Language

A Moscow-based startup called Mostik—Russian for "bridge"—has created a novel approach to AI model interaction that bypasses traditional text output entirely. Rather than having one model feed its textual responses into another, the technique allows different AI systems to exchange information directly through the underlying mathematical parameters that shape their behavior.

How the Bridge Works#

The method leverages the weights that determine how each model converts user input into output. By establishing direct pathways between these numerical structures, Mostik enables smaller and less capable systems to tap into the capabilities of larger ones far more efficiently than conventional methods require. In a demonstration project, the company linked the 753-billion-parameter version of GLM-5.2 with Qwen-3.5, a 4-billion-parameter model designed to run on mobile devices. The resulting hybrid system delivered performance exactly midway between both models while costing only one-twentieth as much as the full GLM deployment.

CEO Sasha Malysheva credits ensemble techniques—the principle that multiple perspectives combined outperform individual expertise. She compares the concept to a well-known statistical finding: when many ordinary people estimate something collectively, their averaged guess often beats a single expert's judgment. Applied to AI, this principle suggests that blending diverse models may prove superior to endlessly scaling up monolithic systems.

Competition and Real-World Impact#

The startup has pushed its approach to tangible success by developing a model that reached the top of ARC-AGI 3, a famously challenging AI competition. The team remains tight-lipped about those specific details pending the contest's conclusion, but the demonstration with open-weight Chinese models illustrates the practical potential.

Industry observers see promise in the approach. Karl Tuyls, a former Google DeepMind researcher, notes that the technique lets organizations "approach large-model quality without the large model handling the entire loop," making it ideal for efficiency-conscious deployment. Vladimir Arustamian, tech lead at AI firm Lovable, was struck by how much Mostik accomplished in months—work he expected would take years.

Implications for AI Development#

Mostik's breakthrough could reshape how the AI industry advances. Rather than betting everything on bigger training datasets and larger models, the approach suggests competitive advantage may lie in orchestrating specialized systems—domain-specific models for biology, physics, and other fields—that work in concert with frontier capabilities.