Russian Startup Mostik Says It Can Make AI Models ‘Talk’ Without Text

A Russian startup called Mostik says it has built a way for AI models to exchange information without generating text, a technique the company says can make small models far more capable and cheaper to run. The method matters because it could help open-weight systems compete more efficiently with proprietary frontier models from companies such as OpenAI and Anthropic.

The idea, developed by a team of Russian mathematicians and AI researchers, uses the internal mathematical structures of models rather than their word outputs. In a demonstration, Mostik says it linked a huge Chinese model with a much smaller mobile-friendly one and produced a hybrid system that cost far less than the largest model while landing midway in performance.

Behind the project is a mix of high-level theory and practical engineering ambition. Mostik’s founders and researchers argue that AI does not have to advance only by making individual models larger. Instead, they believe the next leap may come from letting models collaborate more directly, as if they were passing signals through a hidden bridge rather than chatting in plain language.

What Mostik says it has built

Mostik describes its core innovation as a bridge between models that allows one system’s capabilities to be transferred into another without forcing either model to generate text first. In practical terms, that means the useful knowledge embedded in a larger model can be shared with a smaller one more efficiently than standard model chaining.

Traditional ensembles and multi-model pipelines usually work by passing text outputs from one system into another. That approach can improve accuracy, but it also adds delay, token costs and extra computational overhead. Mostik says its method avoids much of that by working directly with values inside the models’ weights, the numerical parameters that shape how prompts become outputs.

The result, the company says, is a kind of machine-to-machine communication that resembles telepathy more than conversation. The models do not need to explain themselves to each other in words; instead, the bridge aligns internal representations so they can cooperate faster and at lower cost.

Why this matters for AI infrastructure

This matters because the AI industry is under growing pressure to reduce inference costs while improving performance. Frontier models can be expensive to run, especially when they are deployed at scale across consumer products, enterprise tools and autonomous agents. If Mostik’s approach works outside the lab, it could make model deployment significantly more efficient.

It also strengthens the case for open-weight models, which can be downloaded, modified and run by third parties. Those systems often lag behind the biggest closed models in raw capability, but they are cheaper, more flexible and more widely accessible. A method that boosts their performance without forcing them to grow dramatically larger could reshape the competitive landscape.

How does the bridge work?

Mostik says its system connects models through their weights, the mathematical values that determine how each model processes an input. Rather than converting the output of one model into words and then feeding those words into another model, the bridge attempts to transfer information in a more direct numerical form.

The company has not disclosed the full technical details of the method, and that secrecy appears deliberate. Mostik has reportedly been working on the technology partly to gain an edge in a difficult AI evaluation challenge, and it has declined to share some specifics of the system because it wants to remain competitive in the contest.

Even so, the broad concept is clear enough. Instead of treating two models as separate black boxes, Mostik’s approach tries to connect them in a way that lets them cooperate inside the mathematical space where AI systems actually do their work. That could allow a smaller model to inherit some of the strengths of a larger one without copying the entire system.

Sasha Malysheva, Mostik’s chief executive, said the company’s work builds on a familiar machine-learning insight: combining models often outperforms relying on a single one. She also framed the company’s goal as finding a way to do that more efficiently than conventional pipelines.

What is being demonstrated?

To show the concept in action, Mostik created a bridge between two Chinese open-weight models: a very large version of GLM-5.2 and a much smaller Qwen-3.5 model designed to run on mobile hardware. The smaller model has about 4 billion parameters, while the larger one is said to have 753 billion.

According to the company, the hybrid system cost only a fraction of the full-size model to operate. Mostik says the setup ran at roughly one-twentieth the cost of the larger model and delivered performance that sat about halfway between the smaller and larger systems. That kind of result, if reproducible, would be attractive to developers looking for a better cost-to-performance balance.

The test also illustrates the company’s broader pitch: that smaller models can be made much more useful when paired with stronger systems in a way that is computationally smarter than brute-force scaling.

Why are researchers excited about model-to-model communication?

Researchers are excited because the method could make AI systems more modular, more efficient and easier to specialize. Rather than training one model to do everything, developers could combine distinct models for different tasks and connect them in a way that reduces waste.

That would fit a broader trend in AI toward specialization. A general model might still handle broad reasoning, but smaller expert models could focus on biology, physics, code, translation or other fields. If those specialist models can be linked efficiently, the whole stack could become more powerful than the sum of its parts.

It would also create new opportunities for open-weight ecosystems, where developers can experiment with model architectures, deployment strategies and task-specific adaptations without being locked into one proprietary provider’s roadmap.

How does this compare with ordinary ensembles?

It compares favorably in theory because ordinary ensembles usually require one model to produce text and another to read it. Mostik’s method aims to preserve more information and cut out unnecessary translation steps, which could reduce latency and costs.

The company’s CEO argues that model ensembles are already known to outperform single models. The question is not whether combination helps, but how to make combination far more efficient. Mostik’s answer is to move the communication layer deeper into the model itself.

That idea is also where the “bridge” metaphor comes from. Instead of building a long chain of separate model calls, the startup wants to create a direct crossing between internal representations so that one system can benefit from another with less friction.

Who is behind Mostik?

Mostik is led by Sasha Malysheva, who says the concept grew out of the team’s belief that AI progress will not be determined solely by ever-larger models. She argues that the field may eventually value smart composition as much as raw scale.

The startup’s chief scientist is Stanislav Smirnov, a mathematician at the University of Geneva and a Fields Medal winner. His involvement gives the project a strong theoretical pedigree and signals that the company is not just chasing a product idea, but trying to build mathematical foundations for how AI systems can interact.

Other industry figures familiar with the company have described the progress as unusually fast. Some say the team has moved from concept to working system in a matter of months, a timeline that suggests both technical intensity and a practical appetite for rapid iteration.

Smirnov said the central challenge is that two models do not automatically share a common mathematical language. In his view, there is not yet a perfect framework for expressing that relationship, which is why the company’s bridge approach is important.

How the team frames its mission

Malysheva presents the startup as a challenge to the assumption that scale alone will define the future of AI. In her view, the industry may be overcommitted to the idea that bigger is always better, when the real gains could come from combining systems intelligently.

That argument has wider implications. If a bridge can connect a frontier-scale model with a domain-specific model, then the number of useful AI systems could expand dramatically. Biology-focused models, physics-focused models and other specialist tools might become more common if they can be integrated without excessive overhead.

Such a shift would echo how modern software increasingly relies on modular services rather than monolithic applications. In AI, that could mean a more flexible ecosystem where different models cooperate instead of competing for one-size-fits-all dominance.

What could this mean for open-weight models?

It could make open-weight models much more competitive if the technique proves reliable and scalable. Open systems often win on accessibility, transparency and customization, but they can struggle against the most advanced closed models on raw capability.

A bridge-based method could narrow that gap by letting smaller open models absorb part of the strength of larger systems. That would be especially valuable for developers who want to run AI on limited hardware, such as consumer devices, edge systems or cost-sensitive enterprise infrastructure.

It could also shift the economics of AI deployment. If a smaller model can be lifted closer to frontier quality without needing frontier-scale resources, then more companies could afford to build advanced AI features into their products.

• Lower inference costs for apps and services
• Better performance for mobile and edge devices
• More room for specialized open-weight models
• Potentially stronger competition with closed AI systems

Why mathematicians may be central to the next wave of AI

Mathematics may become increasingly important because the hardest AI problems are no longer just about data collection or scaling compute. They are also about understanding how models represent information, how those representations can be aligned and how separate systems can cooperate without losing signal.

Smirnov believes that a deeper mathematical analysis might do more than improve engineering. It could also help explain how AI systems reason and possibly reveal parallels with human cognition. That is a speculative claim, but it reflects a broader ambition in the field: not just to build better systems, but to understand what is happening inside them.

That kind of work is difficult because modern models are famously opaque. They can be effective without being interpretable, which makes bridge-like methods doubly interesting: they may offer both practical gains and a new lens on model internals.

Could this change how AI is researched?

Yes, if the approach proves broadly useful, because it would encourage researchers to think about AI as a network of interoperable components rather than a single giant model. That change would influence training, deployment and evaluation.

It might also create demand for new mathematical tools that describe how representations move between models. If those tools mature, the field could end up with a richer framework for model composition, transfer and collaboration.

Timeline of Mostik’s emergence

Mostik is still a young startup, but its public story has already developed quickly. The company appears to have gone from theory to demonstration in a short time, while keeping some of its most important details under wraps as it pursues competitive AI benchmarks.

What to watch next

The big question is whether Mostik can show that its bridge technique works beyond one carefully chosen demonstration. In AI, many promising ideas look impressive in limited settings but become harder to trust when tested across model families, tasks and deployment environments.

If the company can prove that its approach is robust, the implications could be significant. It could influence how developers architect multi-model systems, how startups build around open-weight infrastructure and how frontier labs think about cost reduction.

If not, it may still be a useful research step. Even partial success could reveal more about how models encode knowledge and how that knowledge can be shared across systems.

For now, Mostik is positioning itself as both a technical outlier and a philosophical challenge to the current AI mainstream. The startup’s message is simple: the future of intelligence may not belong to one huge model alone, but to networks of models that can understand each other in ways humans cannot yet fully describe.

Key facts at a glance

Here is a concise summary of the main details.

For now, Mostik’s approach remains a promising but still closely held bet on a different path for AI progress. Whether it becomes a foundational technique or a clever bridge to nowhere will depend on what happens when other researchers try to cross it.