AI Models Skip Words to Talk Faster — A Fields Medalist Builds the Bridge Between Them

The material of that bridge is model weights — the blocks of numbers that determine how a prompt turns into an output. Mostik finds values inside those numbers that two models can share, and uses them to link the models together. That makes it possible to pass a large model's ability into a smaller model at a fraction of the usual cost.

The company says a model built this way reached the top of ARC-AGI 3, one of the hardest benchmarks in AI. It hasn't disclosed which models it linked or how, citing the contest win as its reason for withholding details.

To put it plainly: until now, the standard way to combine two models was to have one write out its answer in text for the other to read. It's a bit like two coworkers who can only share a thought by writing a full letter and handing it over. What Mostik has done is eliminate the letter and connect the two minds directly with a short bridge.

The cost of writing a letter

Combining multiple models isn't a new idea. Setups where a large model plans and a small model executes, or where several models vote on an answer, are already common. The bottleneck is the channel between them. The standard approach feeds one model's output directly into the next model's input — and that channel carries two kinds of cost.

The first is time and money. Text is generated one token at a time. If the first model takes time to produce 500 tokens, the second model then has to read and understand those same 500 tokens from scratch. Linking two models ends up meaning the computation happens twice.

The second is loss. The computation happening inside a model is a numerical array with thousands of dimensions, and the moment that gets converted into a sentence, most of that array is discarded. A subtle judgment like "I'm not certain, but A seems slightly more likely" gets compressed into a single flat sentence: "The answer is A." The receiving model sees only that flattened sentence and has to guess at the nuance all over again.

This isn't an observation unique to Mostik. A paper published last year, Cache-to-Cache, made the same point: connecting models through text "forces internal representations into token sequences, losing rich semantics and introducing token-by-token latency." The paper proposed directly linking the internal caches of two models instead, and reported that this approach beat text-based linking by 3.1–5.4% in accuracy while running 2.5 times faster on average. Academic research had already been pointing toward the same conclusion — that models do better when they stop talking to each other in words.

The bridge sits on the weights

Mostik chose to build a bit deeper inside the system. Instead of aligning models through the values passed back and forth between them, it aligns them at the level of weights — the numbers that stay fixed once training is finished. Weights are like a map showing how a model divides up its understanding of the world. Because two different models are drawn from different data, they carry different maps, and the same concept can end up plotted at completely different coordinates on each one. Building a bridge means finding a conversion table between those two maps.

Once that conversion table exists, a large model's judgment can be handed to a small model without ever being turned into a sentence. The small model runs on its own, and whenever it hits something it can't solve, it briefly borrows a glance at the large model's map before continuing. It's less like exchanging letters and more like peeking over someone's shoulder.

One-twentieth the cost, exactly in the middle on performance

The model used in the competition is undisclosed, but to illustrate the idea, the company built a demo using two Chinese open-weight models. One was the largest version of GLM-5.2, with 753 billion parameters; the other was Qwen-3.5, a 4-billion-parameter model that runs even on phones. The size gap between them is roughly 188x.

Models linked	Parameters	Where it runs
GLM-5.2 Max	753 billion	Data center
Qwen-3.5 Small	4 billion	Phone
Bridged hybrid	—	Cost: 1/20th of GLM alone · Performance: exactly midway between the two

Read those numbers again. The cost dropped to one-twentieth, but performance didn't sink to the bottom — it landed in the middle. The point of this demo is that pairing in a model 188 times smaller didn't drag performance down toward the small model's level. For anyone running a service, the appeal is obvious: the cost sits near the small model's price tag, while answer quality starts from somewhere in the middle.

Guessing the weight of a pig

Malysheva says this approach is rooted in a very old piece of statistical common sense. "It's well known in machine learning that an ensemble of models outperforms any single model." Inside the company, there's apparently a running joke about guessing a pig's weight: gather guesses from a random crowd, average them, and the crowd beats a single expert at estimating the animal's true weight.

The original version of this story comes from a 1906 livestock fair in Plymouth, England. Statistician Francis Galton collected entries from a contest to guess the weight of an ox. Among 787 valid entries, the median guess was 1,208 pounds, and the ox's actual dressed weight was 1,197 pounds — an error of less than 1%. Individuals were wrong in different directions, so pooling their guesses canceled out the errors. Galton published the result the following year under the title "Vox Populi."

AI models behave similarly. When several models answer the same question, each tends to make different mistakes, so pooling their answers often beats relying on any single model. The catch, until now, was that pooling required every model to write out its full answer. Asking ten models meant receiving ten letters. Mostik's bridge is aimed squarely at eliminating that cost of writing letters.

Why a Fields Medalist is working on this problem

Mostik's chief scientist is Stanislav Smirnov, a professor at the University of Geneva and winner of the 2010 Fields Medal — an award given once every four years, only to mathematicians under 40. There's a reason someone with that background is leading research at a brand-new startup: this is a math problem before it's an engineering problem.

Smirnov says finding common ground between two AI models is surprisingly hard. "It seems like we don't yet have the right mathematical language for this." In other words, there's no existing tool that lets you place two models side by side and say "this part here means the same thing as that part there." That means the bridges built today are more like temporary structures put up before that language exists. "Temporary" might sound like a weakness, but in mathematics, whoever crosses first without the right tools often ends up shaping what the next tool looks like.

He also believes this work could reveal something new about how AI models actually function, and how closely that resembles the human brain. Digging into the conversion table between two models, he suggests, might expose something machines and humans share when working through a hard problem.

No paper yet

It's worth pausing here. What's public right now is a demonstration, not a verified result. The company hasn't explained its core method, citing the competition win as the reason, and there's no paper or benchmark suite yet that outside researchers can run for themselves. The figures — "one-twentieth the cost, exactly midway in performance" — are numbers the company has stated, not numbers a third party has reproduced.

People familiar with the company are largely positive. Vladimir Arustamian, CTO of AI software company Lovable, says: "This team has been at this for just a few months, and they already have working something I expected would take years." He believes that if Mostik can pair frontier models with specialized models in fields like biology or physics, it would open the door to training far more domain-specific models.

Karl Tuyls, a researcher formerly with Google DeepMind, offers a more practical framing. The technology, he says, means "you can get close to large-model quality without running the full process through the large model — a small model running alongside it can deliver a substantial improvement on its own." For anyone who needs to run models as efficiently as possible, he says, there's little reason not to use it.

Still, it's best to keep the order of claims straight. What's solid right now is that this direction looks plausible. How good it actually is remains, for now, the company's own claim. What needs to come next is a paper, and code that others can run themselves.

The scale tips toward open weights

Building this kind of bridge requires being able to look inside the weights of both models. Closed models hidden behind an API can never serve as bridge material in the first place. That means if this approach catches on, the models that gain value are the ones with open weights. It's no accident that the demo used two Chinese open-weight models — right now, the large models freely available to connect and combine are concentrated there.

If this pattern holds, the open-weight camp gains another weapon to compete against closed models from frontier labs like Anthropic or OpenAI. Even if a single open model falls short on its own, a cheaply bridged combination of several could hold its own in a matchup.

That's the picture Malysheva sees too. "Personally, I don't think the future holds a single giant model, or that capability will come from scaling models up." She was referring to the strategy of making models bigger and feeding them more data. In her view, capability is naturally distributed across many models — which means the real skill lies in bringing that distributed capability together.

Reading this from Korea

There are two takeaways worth drawing from this story for a Korean audience.

The first is about cost. The cost of running an AI service ultimately comes down to which size of model gets called, and how often. Until now, there were really only two options: call an expensive large model, or use a cheap small model and accept lower quality. The bridging approach opens up a third option in between — a small model runs by default, and only borrows the large model's judgment briefly when it hits something it can't handle. This is worth watching closely for any team seriously committed to running models on-device.

The second is about materials. Bridges can only be built on top of models with open weights. Choosing a foundation model used to mean weighing performance against price. Going forward, "can this be bridged with other models later" becomes one more item on that list — one more reason open weights count as an asset.

Malysheva says she only realized she had a gift for math after her older brother told her she'd never be able to solve the math olympiad problems he was working on. A few years later, she was studying at one of St. Petersburg's top schools. She says she heard similar things when she set out to build this bridging approach. "People kept saying it would be too hard for a young girl to pull off. I decided I needed to prove them wrong."

For the past several years, AI progress has been summed up in a single phrase: bigger and more. The belief that scaling up models and feeding them more data would keep pushing performance higher has driven investment, power grids, and data centers alike. What Mostik is pushing against is the flip side of that belief. If capability is already scattered across many models, the next challenge isn't building one bigger model — it's cheaply connecting the pieces that already exist.

And in that work of connecting pieces, the single most expensive component turned out to be words themselves. The one step of converting internal computation into human-readable sentences was quietly consuming time, money, and information all at once. Once you accept that models don't need humans to understand what they're saying to each other, that step disappears. Whether the bridge these Russian mathematicians built is truly solid won't be clear until a paper appears. But the broader direction — that skipping words altogether works better — is already something multiple lines of research are pointing toward. The next round in AI may not be decided by who builds the biggest model, but by what happens in the narrow channel between one model and another.