A Russian startup called Mostik has built a way for AI models to exchange information directly through the mathematical values in their weights, bypassing the text outputs that ensemble systems normally pass between models. In a public demonstration, Mostik bridged the 753-billion-parameter GLM-5.2 with a 4-billion-parameter version of Qwen-3.5, producing a hybrid system that costs one-twentieth of running the full GLM model while landing exactly halfway between the two on performance. The company has also used the technique to build a separate model that has climbed to the top of ARC-AGI 3, one of the harder public benchmarks for reasoning-capable systems.
Mostik — Russian for bridge — is led by CEO Sasha Malysheva, who developed the approach, with 2010 Fields Medalist Stanislav Smirnov of the University of Geneva as chief scientist. The pitch is that a small on-device model can absorb the capability of a much larger frontier-scale model without paying the frontier-scale inference bill. A 4-billion-parameter model that can run on a phone, plugged into a 753-billion-parameter model through the Mostik bridge, ends up performing at roughly the midpoint of the two — a substantial jump for the smaller model at a fraction of the compute needed to serve the larger one end-to-end.
Ensembling AI models is not new. What is new is the interface. Standard ensembles feed the text output of one model into the input of another, which is slow and expensive because every hop pays for a full generation pass. Mostik claims to have found a way to let the models communicate at the level of their weights, which are the numerical parameters that translate an input prompt into an output. The approach is essentially a translation layer between two different models' internal representations, sidestepping the natural-language round trip entirely.
“It's well-known in machine learning that ensembles of models perform better than individual ones.”— Sasha Malysheva, CEO of Mostik
Key facts
- 01Mostik's hybrid pairing of GLM-5.2 (753B parameters) with Qwen-3.5 (4B parameters) runs at one-twentieth the cost of the full GLM model.
- 02The combined system's performance lands exactly halfway between the two source models.
- 03A separate Mostik-built model has topped the ARC-AGI 3 leaderboard, though the team is withholding details ahead of the contest.
- 04Chief scientist Stanislav Smirnov, a 2010 Fields Medalist and University of Geneva professor, leads the mathematical work.
- 05The technique lets models exchange information via their weights directly, skipping the text-output round trip that ensembles normally require.
Malysheva frames the future of AI as closer to guessing the weight of a pig than to a single dominant model — a running joke inside the company, referencing the statistical result that averaged guesses from a crowd tend to beat any individual expert. Combining outputs from many models, in her view, is a more efficient path to capability than making one model progressively larger. She does not believe the field will end up with a single monolithic model, and does not think future capability will come from scaling parameters and data alone.
If the technique holds up, the commercial implication for open-weight models is direct. The two models Mostik bridged in its demo, GLM-5.2 and Qwen-3.5, are both Chinese open-weight releases. Pairing them via a bridge produces a system whose economics look nothing like calling a closed frontier API. That is the competitive angle: open-weight models, until now judged mainly on their standalone benchmark scores, become far more attractive when they can be combined cheaply into hybrid stacks that rival proprietary offerings from Anthropic and OpenAI.
Vladimir Arustamian, the tech lead at AI software company Lovable, said the approach could unlock a wave of specialized model training if frontier models can be paired with domain-specific ones in biology, physics, and similar fields. Karl Tuyls, a former Google DeepMind computer scientist familiar with Mostik's work, said the method lets operators approach large-model quality without routing the entire workload through the large model, and called it a no-brainer for anyone running inference at scale.
“This team has been at it for a matter of months and already has something running that I would have guessed was years out.”— Vladimir Arustamian, Tech lead at Lovable
The efficiency case matters because inference cost is now the dominant line item for anyone deploying large models in production. If a 4-billion-parameter model paired with a 753-billion-parameter bridge partner delivers midpoint quality at one-twentieth the compute of the full 753B model, the operating math changes for every team that has been forced to choose between capability and unit economics. That trade-off has driven a wave of smaller-model releases from every major lab over the past year.
There are open questions. Smirnov acknowledged that finding common ground between two models' internal representations is mathematically hard, and that there is no established formal language for it yet. The demonstration numbers — one-twentieth the cost, midpoint performance — are Mostik's own, on a specific pair of open-weight models, and have not been independently replicated across the diverse pairings that would prove the technique generalizes. The ARC-AGI 3 result is impressive on its face, but Mostik is withholding details until the competition concludes, which limits outside scrutiny.
“There seems to be no appropriate mathematical language yet.”— Stanislav Smirnov, Mostik chief scientist and 2010 Fields Medalist
The bet Mostik is placing is that the next capability gains in AI come from composition rather than scale. If bridges between models work as advertised, the value of every open-weight release rises, because each new model becomes a component that plugs into a larger system rather than a standalone product competing head-to-head with GPT-class systems. That would shift some of the leverage in the AI market away from the labs training the biggest models and toward the teams building the connective tissue between them — a structural change worth watching, whether or not Mostik itself ends up being the company that captures it.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.
