Mostik lets AI models talk to each other through their weights

Mostik, a startup founded by Russian mathematicians, has built a method that lets separate AI models communicate through the numerical values in their weights rather than by exchanging text. The company's CEO, Sasha Malysheva, developed the approach; the startup takes its name from the Russian word for bridge, a nod to how it connects models to one another. In practice, the technique lets the capabilities of a larger model be transferred to a smaller one far more efficiently than the usual method of feeding one model's output into another.

Mostik has already used the technique to build a model that reached the top of ARC-AGI 3, a notoriously hard AI benchmark competition; the company would not share further details because it wants to win the contest. To show how the underlying method works, the team instead built a public demonstration: a bridge between the largest version of GLM-5.2, a Chinese open-weight model with 753 billion parameters, and a 4-billion-parameter version of Qwen-3.5 small enough to run on a mobile device. The resulting hybrid system costs one-twentieth of running the full GLM-5.2 model, and its performance lands exactly halfway between the two source models.

Malysheva frames the idea around a well-known result in machine learning: ensembles of models tend to outperform any single model on its own. She compares it to the classic finding that a crowd's averaged guesses at a pig's weight beat an individual expert's estimate. The usual way to combine AI models, feeding one model's text output into another, costs significant time and money. Mostik's team found a way for models to exchange information without producing that intermediate text at all. Malysheva says she personally doubts AI progress will keep coming from a single monolithic model scaled up with more data; she thinks combining many different models may prove the better path, and if Mostik's approach catches on, it could raise the competitive value of open-weight models against closed, proprietary systems from labs like Anthropic and OpenAI.

Two people outside the company vouch for the technique. Vladimir Arustamian, tech lead at the AI software company Lovable, who knows the Mostik team, says that if the approach lets frontier models pair with domain-specific ones in fields like biology or physics, it would push many more specialized models to get trained; he adds that the team has been at the work only a matter of months yet already has something running he would have guessed was years away. Karl Tuyls, a former Google DeepMind computer scientist familiar with Mostik's technology, says the method lets a smaller model approach large-model quality without the large model handling the full workload, and calls it a no-brainer for anyone trying to run models as efficiently as possible.

Mostik's chief scientist is Stanislav Smirnov, a professor at the University of Geneva and a 2010 Fields Medalist. He says finding common ground between two AI models is surprisingly hard, and that no appropriate mathematical language for it exists yet; Mostik's bridging approach is a way to close that gap in the meantime. Smirnov also suggests the work could eventually reveal something about how AI models actually function and how that compares to human reasoning, since a deeper mathematical analysis might turn up commonalities between how AI systems and people work through difficult problems.

Malysheva says she found her own talent for math after her older brother told her she would not be able to solve the Math Olympiad problems he was studying; she went on to study at one of the top schools in St. Petersburg. More recently, some peers warned her the weight-bridging approach would be too hard to pull off. "They said it might be too hard for a young girl," she says. "I decided I need to prove them wrong."

Key facts

Mostik, a startup founded by Russian mathematicians, built a technique that lets AI models exchange information through the values in their weights instead of producing text output.
In a public demo, the team bridged the 753-billion-parameter GLM-5.2 with a 4-billion-parameter version of Qwen-3.5 that runs on a mobile device; the resulting hybrid costs one-twentieth of the full GLM-5.2 and performs exactly halfway between the two.
Mostik says it used the same underlying technique to build a model that reached the top of the ARC-AGI 3 benchmark competition, but withheld details because it wants to win the contest.
Mostik's chief scientist is Stanislav Smirnov, a University of Geneva professor and 2010 Fields Medalist, who says no adequate mathematical language yet exists for finding common ground between AI models.
Outside observers, including a former Google DeepMind scientist and the tech lead at Lovable, describe the approach as a practical way to run models more efficiently and to pair frontier models with smaller, domain-specific ones.

Why it matters

Combining several AI models usually beats relying on one, but the standard way to do it, feeding one model's text output into another, is slow and expensive. Mostik's pitch is that models can share what they know directly through their weights instead, skipping the text step. If the technique holds up outside the one demonstrated case, it could make it cheaper to boost a small model with a larger one's capability, and Malysheva argues it could make open-weight models more competitive with closed systems from labs like Anthropic and OpenAI, since open weights become more valuable when they can be recombined this way.

Who it affects

Anyone running open-weight models who wants better results without paying for a full-sized model would benefit most directly, since the demonstrated hybrid cuts cost to a twentieth of the large model while keeping mid-tier performance. Arustamian's comment points at a further audience: teams that want to pair a frontier general model with smaller domain-specific ones, in fields such as biology or physics, rather than training one model to do everything. Mostik itself is also chasing benchmark competitions like ARC-AGI 3, so AI research groups tracking that leaderboard are a third audience.

How to use it

The source gives no pricing, release date or public access details for Mostik's technique; it is described as a company capability demonstrated to a reporter, not a shipped product, so nothing here should be read as an offer. The only concrete usage detail is the demo itself: pairing a 753-billion-parameter GLM-5.2 with a 4-billion-parameter Qwen-3.5 that is small enough to run on a phone.

How solid is it

The claims come from a single interview with Mostik's CEO and chief scientist, plus supporting comments from two outside figures who know the team, one a former Google DeepMind scientist and one Lovable's tech lead; neither is described as having independently tested the system. Mostik has not published a paper or explained the underlying mechanism. Smirnov, the chief scientist, says outright that no proper mathematical language yet exists for what the team is doing, which is itself a sign the method has not been formalized. The ARC-AGI 3 result is asserted but undocumented: Mostik declined to give details because it wants to win the contest, so that claim cannot currently be checked. The GLM-5.2/Qwen-3.5 numbers, cost and performance figures come from the company's own account of its demonstration, not an outside benchmark.

Risks and caveats

Every figure in this story, the 753 billion and 4 billion parameter counts, the one-twentieth cost, the halfway performance, and the ARC-AGI 3 ranking, traces back to what Mostik told the reporter, with no published paper, benchmark scores or reproducible details behind it. No funding, headcount, founding date or public release plan is disclosed. Readers should treat this as an early, unverified account of a promising-sounding technique rather than a validated result.

“It's well-known in machine learning that ensembles of models perform better than individual ones”

— Sasha Malysheva, Mostik's CEO