Mostik, a startup founded by Russian mathematicians, says it has found a way for different AI models to collaborate through their internal mathematical representations instead of passing text back and forth.
The company’s name is the Russian word for bridge, reflecting an approach that connects the values embedded in model weights. Mostik says this allows capabilities from a large model to improve the performance of a smaller one without requiring the larger system to handle the entire task.
In a demonstration reported by Wired, the startup connected two Chinese open-weight models: a 753-billion-parameter version of GLM-5.2 and a 4-billion-parameter Qwen-3.5 model capable of running on a mobile device. The hybrid reportedly cost one-twentieth as much as the full GLM model while delivering performance exactly halfway between the two.
A mathematical shortcut around text generation
Combining multiple models is not a new idea. Model ensembles can improve results by pooling different systems’ outputs, much as averaging several independent estimates can outperform a single expert judgment. Most existing arrangements, however, require one model to produce an answer that another model then reads and processes.
That sequence consumes time and computing resources. Mostik CEO Sasha Malysheva says her team’s method lets models communicate before producing text, working instead with the numerical structures that shape how prompts become outputs.
The company has not published enough technical detail in Wired’s report to independently evaluate the mechanism. It is therefore not possible from the source to determine how widely the bridge works across architectures, how it behaves on different tasks or whether the reported cost relationship holds outside the demonstration.
Still, the result presents a concrete alternative to running a frontier-scale model through every stage of a workload. Karl Tuyls, a former Google DeepMind computer scientist familiar with the technology, said the method can approach large-model quality by adding a smaller model alongside it.
Open models could gain more specialized value
Mostik argues that its approach could make open-weight models more competitive with closed systems from companies such as Anthropic and OpenAI. If a smaller model can absorb useful capabilities from a larger partner during execution, organizations may gain more flexibility in how they balance performance and cost.
Vladimir Arustamian, technical lead at AI software company Lovable and an acquaintance of the team, suggested that bridges between frontier models and domain-specific systems could encourage more specialized training in fields such as biology and physics. That is a projection rather than an outcome demonstrated in the article.
Malysheva also questions whether progress must continue through increasingly monolithic models trained with more data. She sees combinations of distinct systems as a possible alternative. The startup’s method would need to show that such combinations remain stable and useful at scale before that view can be tested against the industry’s prevailing strategy.
Mostik says it has already used the approach to build a model that reached the top of ARC-AGI 3, a difficult AI competition. The team withheld further information because it wants to win, leaving the claim without enough disclosed evidence for outside assessment.
Bridging models remains an unsolved language problem
Mostik’s chief scientist is Stanislav Smirnov, a University of Geneva professor and 2010 Fields Medalist. He said that finding common ground between two AI models is unexpectedly difficult because there is not yet an adequate mathematical language for the task.
The bridge is the company’s working answer to that gap. Smirnov believes further mathematical analysis could also reveal similarities between how models and humans reason through difficult problems, although Wired reports this as a possibility rather than a research conclusion.
Mostik’s early examples combine ambitious claims with limited disclosure. The GLM–Qwen demonstration offers specific figures on cost and relative performance, while the ARC-AGI result and broader implications remain largely unexamined in public.
The immediate idea is nevertheless clear: AI systems do not necessarily have to explain themselves to one another in words. If their internal representations can be connected reliably, a smaller system may borrow useful capacity from a larger one without paying the full computational price each time.
