Mostik, a startup whose name comes from the Russian word for a small bridge, says it has found a way for two AI models to exchange information directly through the mathematical values held in their weights, with neither one writing a word of text. As a demonstration, it joined two Chinese open models: GLM-5.2, the largest version at 753 billion parameters, and Qwen-3.5, a 4-billion-parameter build small enough to run on a phone. The hybrid costs 20 times less than the full GLM, and its performance sits exactly midway between the two. The company says it has also used the method to build a model that shot to the top of ARC-AGI 3, one of the harder contests in the field.
The underlying observation is old. Sasha Malysheva, Mostik's CEO, points to ensembles: in machine learning it has long been known that several models together often beat one. The company explains it with a joke about guessing the weight of a pig — the example from mathematical folklore in which a crowd of ordinary people, their guesses averaged, estimates the weight more accurately than a single expert. Combining the answers of several AI models works much the same way.
What has made ensembles awkward in practice is the plumbing. The usual arrangement has one model produce an answer and hand it to another, which costs real time and real money at every hop. Mostik's claim is that the handoff does not need to happen in words at all. The weights are what turn a query into an answer, so the exchange happens there.
Read as an engineering result rather than a scientific one, the demonstration is easy to underrate and easy to oversell. Halfway between a 753-billion-parameter model and a 4-billion-parameter one, at a twentieth of the large model's cost, is a genuinely good trade for anyone running inference at volume. It is also, by the company's own description, an interpolation. Nothing in the demo shows the pair doing something neither model could do; it shows a cheaper point on a curve that already existed.
Karl Tuyls, a former Google DeepMind computer scientist familiar with the technology, describes the appeal in exactly those terms: it gets close to the quality of a large model without making that model chew through the whole task, with a smaller one running alongside to supply a noticeable improvement. For anyone who needs to run models as efficiently as possible, he says, the method is the obvious choice.
Vladimir Arustamyan, chief technology officer at the AI software company Lovable and an acquaintance of the team, sees a different consequence. If the technique can connect frontier models to specialised ones — for biology, say, or physics — more of those specialised models will get trained, because there will finally be a cheap way to put them to work. He notes that the team has been at this only a few months and has already shipped something that looks like the product of years.
Mostik's chief scientist is Stanislav Smirnov, a professor at the University of Geneva and a Fields medallist. He says the hard part was not what anyone expected: finding a common language between two AI models turned out to be unexpectedly difficult, because no suitable mathematical language for it exists yet. Mostik's method, in his description, is a temporary patch over that gap — which is an unusually honest thing for a company's chief scientist to say about its core product. He argues the work may end up explaining how these models are built at all, and where their reasoning resembles the brain's; a deeper mathematical treatment could expose common patterns in how people and machines work through hard problems.
The most load-bearing claim here is also the one with the least evidence attached. The ARC-AGI 3 result is what would make this more than an inference-cost story, and the team will not describe how it was achieved, because it wants to win the competition. That is a defensible commercial decision and an unsatisfying epistemic one. Everything checkable is the GLM–Qwen demo, where the numbers are pleasant and the result is bounded by construction.
The quiet part is what the method needs to work. If models talk through their weights, someone has to have the weights — which confines the technique to open models and rules out the frontier systems from Anthropic and OpenAI that Malysheva positions it against. That is not a footnote; it is the entire strategic shape of the company. Mostik makes open weights more valuable precisely because they are open, and a wave of adoption would be an argument for releasing weights, not against it.
Malysheva's broader bet follows from that: a future in which capability comes from combining many different models rather than from one monolith, and in which new abilities do not arrive solely through scale — bigger models, more data. She has been told before that something was beyond her. She found she had a gift for mathematics after an older brother told her she would not be able to solve the olympiad problems he was working on; a few years later she was at one of the best schools in St. Petersburg. Colleagues warned her the bridge idea would be too hard to build, and that it might be too hard for a young woman in particular. She set out to prove otherwise.
Which leaves the interesting tension. If weight-level exchange works as well as Mostik suggests, the cheapest path to frontier-grade output stops running through the biggest model and starts running through combinations of mid-sized ones — and the labs best positioned to be combined are the ones that publish their weights, most of them Chinese. The companies that would benefit most from Mostik's bridge are the ones that decided not to keep their side of the river private.