A Russian startup called Mostik has developed a method for artificial intelligence models to exchange information directly through their internal weights, bypassing the need for text generation.
The team works at Mostik, a name chosen for its meaning of bridge. Their technique allows different systems to interact using the mathematical values stored in their parameters. This process transfers the capabilities of a larger model into a smaller one, increasing intelligence without requiring the larger system to process the entire task.
Performance and cost
The startup applied this method to create a model that reached the top position in ARC-AGI 3, a competition known for its difficulty. The developers declined to share further details regarding this specific result to protect their competitive edge.
For a public demonstration, they connected two Chinese open-weight models. One was the 753-billion-parameter version of GLM-5.2. The other was a 4-billion-parameter version of Qwen-3.5 capable of running on a mobile device. The hybrid system costs one-twentieth of the full GLM model while delivering performance exactly halfway between the two.
How it works
Sasha Malysheva, the CEO of Mostik, noted that ensembles of models typically outperform individual ones. She explained this by referencing a common joke within the company about estimating the weight of a pig. A group of random people often guesses the weight more accurately than a single expert when their individual estimates are averaged together.
Combining the outputs of several AI models usually yields better results. Standard practice involves feeding the output of one model into another, a process that consumes significant time and money. Mostik’s approach allows systems to communicate without producing text. If the technology succeeds, it could increase the value of open-weight models and help them compete with closed, proprietary systems from companies like Anthropic and OpenAI.
Future direction
Malysheva suggests that combining various models may be a superior path for advancing AI compared to scaling up monolithic systems. She stated she does not expect future capabilities to come solely from making models larger or feeding them more data.
Vladimir Arustamian, the tech lead at the AI software company Lovable, believes the technology could pair frontier models with domain-specific ones in fields like biology and physics. This would allow for the training of many more specialized models. Arustamian noted the team has been working on this for months and has achieved results he previously thought would take years.
Karl Tuyls, a former computer scientist at Google DeepMind, described the method as a no-brainer for anyone needing to run models efficiently. He explained that large-model quality can be approached without the large model handling the entire loop, providing substantial improvements with a smaller model running alongside.
Mathematical challenges
Stanislav Smirnov, a professor at the University of Geneva and a 2010 Fields Medalist, serves as Mostik’s chief scientist. He stated that finding common ground between two AI models is surprisingly difficult. Smirnov noted there currently seems to be no appropriate mathematical language for this purpose. The current approach acts as a literal bridge across that gap.
Smirnov added that the work might reveal new insights into how AI models function compared to the human brain. A deeper mathematical analysis could show commonalities in how both systems reason over difficult problems.
Origins
Malysheva discovered her talent for mathematics after her older brother told her she could not solve the Math Olympiad problems he was studying. She later studied at a top school in St. Petersburg. Peers recently warned her that the bridge approach would be too difficult to execute. She responded by deciding she needed to prove them wrong.
Malysheva said they told her it might be too hard for a young girl.
