A Russian startup called Mostik has built a mathematical bridge that lets AI models exchange information directly through their internal weights, bypassing the expensive text-based communication that ensemble systems normally require. In a public demonstration, the company fused a 753-billion-parameter GLM-5.2 model with a 4-billion-parameter Qwen-3.5, producing a hybrid system that costs one-twentieth of running the full large model while delivering performance exactly halfway between the two.

How Does This Weight-Sharing Technique Actually Work?

Traditional AI ensembles work by feeding the text output of one model into another, which is slow and expensive because every handoff requires a full generation pass. Mostik's innovation sidesteps this entirely by creating what amounts to a translation layer between two different models' internal representations. Instead of models talking to each other through language, they communicate at the level of their weights, the numerical parameters that translate an input prompt into an output.

The company is led by CEO Sasha Malysheva, who developed the approach, with 2010 Fields Medalist Stanislav Smirnov of the University of Geneva serving as chief scientist. The pitch is straightforward: a small on-device model, like one that runs on a phone, can absorb the capability of a much larger frontier-scale model without paying the frontier-scale inference bill. A 4-billion-parameter model 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.

"It's well-known in machine learning that ensembles of models perform better than individual ones," said Sasha Malysheva, CEO of Mostik.

Sasha Malysheva, CEO of Mostik

Why Should Companies Care About This Development?

The efficiency case matters because inference cost, the expense of running a model after it's been trained, 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 753-billion 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.

The two models Mostik bridged in its demonstration, GLM-5.2 and Qwen-3.5, are both Chinese open-weight releases, meaning their code and weights are publicly available. Pairing them via a bridge produces a system whose economics look nothing like calling a closed frontier API from companies like OpenAI or Anthropic. 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.

What Are the Practical Implications for AI Deployment?

• Cost Reduction: Running a hybrid system costs one-twentieth of the full large model while maintaining performance halfway between the two source models, dramatically improving unit economics for production deployments.
• Specialized Model Pairing: The technique could unlock a wave of specialized model training if frontier models can be paired with domain-specific ones in biology, physics, and similar fields, according to industry observers.
• Inference Efficiency: The method lets operators approach large-model quality without routing the entire workload through the large model, a critical advantage for anyone running inference at scale.

Vladimir Arustamian, 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," said Vladimir Arustamian, tech lead at Lovable.

Vladimir Arustamian, Tech Lead at Lovable

What Questions Remain About the Technology?

There are significant open questions about whether Mostik's approach will generalize beyond the specific demonstration. 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 and 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.

"There seems to be no appropriate mathematical language yet," acknowledged Stanislav Smirnov, Mostik chief scientist and 2010 Fields Medalist.

Stanislav Smirnov, Chief Scientist at Mostik

Mostik has also built a separate model that has climbed to the top of ARC-AGI 3, one of the harder public benchmarks for reasoning-capable systems. However, the team is withholding details until the competition concludes, which limits outside scrutiny of how the weight-sharing technique contributed to that result.

What Does This Mean for the Future of AI Architecture?

Malysheva frames the future of AI as closer to what she calls the "wisdom of crowds" principle, a running joke inside the company that references 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.

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.