The Future of AI Communication: How Startup ‘Mostik’ Is Enabling Machine Telepathy
In a breakthrough that challenges the industry’s obsession with ever-larger models, a new startup called Mostik is rethinking how artificial intelligence interacts. By developing a method for AI models to communicate through their core mathematical “weights”—effectively bypassing the need for slow, traditional text output—the team has created a form of “machine telepathy” that could fundamentally alter the landscape of AI efficiency.
The name “Mostik,” the Russian word for “bridge,” serves as a direct nod to the company’s core innovation: building direct, high-speed conduits between disparate models.
Breaking the Scaling Paradigm
For years, the gold standard in artificial intelligence has been “scaling”—building massive, monolithic models fed on gargantuan amounts of data. However, Sasha Malysheva, CEO of Mostik, believes the industry is heading in the wrong direction.
“I personally do not think we will have a monolithic model in the future, or that the capabilities of models will come from scaling,” Malysheva told reporters recently. Instead, she draws a parallel to the “wisdom of the crowd” theory, where an averaged guess from a group of people is often more accurate than the opinion of a single expert. By allowing models to “talk” to each other directly at the mathematical layer, Mostik aims to leverage the combined intelligence of multiple smaller, specialized models rather than relying on one massive, expensive system.
Engineering a New Kind of Bridge
The practical results of this approach are already turning heads. Mostik recently demonstrated a hybrid system pairing the 753-billion-parameter GLM-5.2 model with a mobile-ready, 4-billion-parameter version of Qwen-3.5. The resulting system costs roughly one-twentieth of the full GLM model while delivering performance exactly halfway between the two—a staggering jump in efficiency.
The startup has also applied its technique to achieve high rankings in the ARC-AGI competition, a benchmark known for testing the reasoning capabilities of AI models. Because they aim to remain competitive in the contest, the team is keeping the finer technical details under wraps.
Industry Impact
Experts familiar with the technology suggest that Mostik’s approach could be a game-changer for the open-source community. Currently, frontier labs like OpenAI and Anthropic dominate the field with massive, proprietary models. By allowing smaller, open-weight models to communicate and combine their capabilities, Mostik could provide a viable path for developers to achieve elite-level performance without the prohibitive cost of running gargantuan infrastructure.
“This team has been at it for a matter of months and already has something running that I would have guessed was years out,” says Vladimir Arustamian, tech lead at the AI software company Lovable.
Karl Tuyls, a former Google DeepMind scientist, notes that the technique is a “no-brainer” for those tasked with optimizing AI operations. “You can approach large-model quality without the large model handling the entire loop,” Tuyls explains.
The Mathematical Frontier
Leading the scientific effort is Stanislav Smirnov, a University of Geneva professor and 2010 Fields Medalist. Smirnov admits that the work is in its early stages, noting that establishing a common mathematical language between different AI architectures is an immense challenge. “There seems to be no appropriate mathematical language yet,” he says.
For now, Mostik is providing the necessary bridge to navigate that void. If the startup succeeds in scaling this technique, it may finally break the industry’s reliance on massive, power-hungry models, ushering in an era of nimble, hyper-efficient AI that can operate on mobile devices while punching well above its weight class.
