A group of Russian mathematicians working for the startup Mostik developed a technique that allows artificial intelligence models to interact without using words. Mostik translates to bridge in Russian. This name reflects the startup approach, which lets different models communicate using the mathematical values found in their weights.
Weights are the internal components that determine how a prompt turns into an output. Through this method, the capabilities of a larger model transfer to a smaller model to increase its intelligence more efficiently. Mostik used this strategy to build a model that reached the top of ARC-AGI 3, a difficult competition for AI models. The team withheld further details to protect their chances of winning the contest.
To demonstrate the technology, the startup created a bridge between two Chinese open-weight models. These included the largest version of GLM-5.2 with 753 billion parameters and a 4-billion-parameter version of Qwen-3.5 designed to run on mobile devices. The resulting hybrid system costs one-twentieth of the full GLM model, and its performance lands halfway between the two individual systems.
Sasha Malysheva serves as Mostik CEO and developed the approach. She shared a company joke comparing the future of AI to guessing the weight of a pig. In mathematics, a group of random people can accurately estimate a pig weight by averaging their combined guesses better than a single expert. Combining the outputs of multiple AI models yields better results in a similar manner.
Traditional model combining requires feeding the output of one model into another, which consumes time and money. The Mostik team figured out a way for models to talk to one another without producing text output. If successful, the method could increase the value of open-weight models and help them compete with proprietary models from frontier labs like OpenAI and Anthropic.
Malysheva believes combining many different models could prove to be a superior way to advance AI. She noted that she does not expect future AI to rely on a single monolithic model or scaling through larger sizes and more data.
Vladimir Arustamian is the tech lead at the AI software company Lovable and knows the Mostik team. He stated that pairing frontier models with domain-specific models in biology and physics could encourage the training of many specialized models. Arustamian added that the team built a functioning system in months that he previously assumed was years away.
Karl Tuyls is a former computer scientist at Google DeepMind who knows Mostik technology. He noted that the technique allows developers to approach large-model quality without a massive model handling the entire loop. Tuyls called the method a logical choice for efficient model operation.
Stanislav Smirnov is a professor at the University of Geneva, a 2010 Fields Medalist, and the chief scientist at Mostik. He explained that finding common ground between two AI models remains difficult because an appropriate mathematical language does not yet exist. He views the Mostik approach as a practical way to bridge the gap.
Smirnov also suggested that the work could reveal new insights into how AI models function and how their operations compare to the human brain. He believes deeper mathematical analysis might show a commonality in how artificial systems and humans reason over complex problems.
Malysheva discovered her talent for math after her older brother claimed she could not solve Math Olympiad problems he studied. She later attended a top school in St. Petersburg. When peers recently warned her that the bridge approach was too difficult, she decided to prove them wrong.
The startup continues to guard details about its ARC-AGI 3 competition entry.
