A Russian Maths Startup Found a Way for AI Models to Share Intelligence Without Swapping Words

Key points

• Mostik, a Russian AI startup, built a method that lets two AI models share abilities directly through their internal maths, without either model generating text output.
• A hybrid system combining a 753-billion-parameter model and a 4-billion-parameter mobile model costs one-twentieth of running the large model alone.
• The resulting hybrid performs exactly halfway between the two models on benchmarks.
• Mostik's approach has placed the team at the top of ARC-AGI 3, one of the hardest AI performance competitions running today.
• Fields Medal-winning mathematician Stanislav Smirnov serves as Mostik's chief scientist.

Inside every AI model is a vast grid of numbers called weights, the values that determine how the model turns your question into an answer. Mostik, a startup whose name is the Russian word for bridge, found a way to read those numbers directly and use them to transfer the skills of one model to another. No words exchanged. No text generated. Just maths talking to maths.

The practical payoff is striking. The team built a bridge between two Chinese open-weight models, meaning models whose internal workings are publicly available. One was the largest version of GLM-5.2, a giant with 753 billion parameters (each parameter is a single adjustable number the model learned during training). The other was a compact Qwen-3.5 model with just 4 billion parameters, small enough to run on a mobile phone.

The hybrid costs one-twentieth as much to run as the big model alone. Its performance lands exactly halfway between the two.

Why does this matter to anyone outside AI research?

Running a powerful AI model is expensive, which is why most capable systems sit behind paywalls or corporate firewalls. Mostik's method could let a small, cheap model borrow muscle from a large one without either side paying the full bill.

"You can approach large-model quality without the large model handling the entire loop, giving you substantial improvements with just a smaller model running alongside," said Karl Tuyls, a former computer scientist at Google DeepMind who reviewed the company's work.

For ordinary users, the downstream effect could be better AI tools on cheaper hardware, including phones.

The standard way to combine AI models today is to feed the output of one into the input of another, a process Wired AI first reported on that chews through time and money. Mostik skips that step entirely by working at the level of the weights themselves.

"It's well-known in machine learning that ensembles of models perform better than individual ones," CEO Sasha Malysheva said.

She draws an analogy from statistics: a crowd of random people guessing a pig's weight will, on average, beat a single expert if you average their guesses. Combine enough AI models and you get a similar lift.

What happens next?

Mostik is currently competing at the top of ARC-AGI 3, a notoriously hard benchmark for AI systems, though the team declined to share details until the contest closes.

Longer term, Malysheva believes the future of AI is not one giant model trained on ever more data, but many smaller specialised models working together. "I personally do not think we will have a monolithic model or that the capabilities of models will come from scaling," she said.

Chief scientist Stanislav Smirnov, a University of Geneva professor who won the Fields Medal, the highest honour in mathematics, in 2010, is candid about the difficulty. "There seems to be no appropriate mathematical language yet," he said. He also believes a deeper analysis of how models share information may reveal similarities to how human brains solve hard problems.

Malysheva started the company after peers told her the bridge approach was probably too hard. "They said it might be too hard for a young girl," she said. "I decided I need to prove them wrong."