What happened?

WIRED's Will Knight reports that a Russian startup called Mostik has built a way for AI models to talk to each other without producing any text. The name is Russian for "little bridge," which is exactly what the approach does: two models interact through the mathematical values in their weights rather than through outputs.

In practice that means the capabilities of a large model can be fed into a small one far more efficiently than they are today. The usual method pipes one model's text output into another, which costs both time and money. Mostik skips that intermediate step.

The demo: a bridge between 753 billion and 4 billion

The team says it used the method to build a model that shot to the top of ARC-AGI 3, a notoriously difficult competition, but would not give details because they want to win it. To demonstrate the idea instead, they bridged two Chinese open-weight models:

  • The largest version of GLM-5.2 — 753 billion parameters.
  • A 4-billion-parameter version of Qwen-3.5 that runs on a phone.
  • The resulting hybrid system costs one-twentieth of the full GLM model.
  • Its performance sits exactly halfway between the two.

Why it matters

Mostik CEO Sasha Malysheva grounds it in a simple observation: it is well known in machine learning that ensembles of models outperform individual ones. Or, as the company's running joke has it, the future of AI resembles guessing the weight of a pig — a handful of random people, averaged, estimate it more accurately than a single expert.

Malysheva draws a bolder conclusion from that: "I personally do not think we will have a monolithic model in the future, or that the capabilities of models will come from scaling." If the method holds up, it could raise the value of open-weight models and make competing with closed frontier models easier.

Vladimir Arustamian, tech lead at Lovable and an acquaintance of the team, says that if Mostik makes it possible to pair frontier models with domain-specific ones in biology or physics, many more specialised models would get trained: "This team has been at it for a matter of months and already has something running that I would have guessed was years out."

How verifiable is it?

This is where caution is warranted. There is no peer-reviewed publication; no detail is shared about the ARC-AGI 3 result, and the only concrete demonstration is a two-model example. "Performance exactly halfway, cost one-twentieth" is a striking sentence, but which tasks it was measured on, and by what benchmark, is not stated.

Karl Tuyls, a former Google DeepMind computer scientist familiar with the technology, sums the method up this way: "You can approach large-model quality without the large model handling the entire loop." Mostik's chief scientist, University of Geneva professor and 2010 Fields medallist Stanislav Smirnov, concedes the difficulty: finding common ground between two models is surprisingly hard because "there seems to be no appropriate mathematical language yet."

What's next?

Smirnov thinks a deeper mathematical analysis could reveal new things about how models actually work and how that compares with the human brain. The company says it will not share its ARC-AGI 3 results until the competition ends — which means independent testing of the method largely waits on that too.