A little-known Russian startup called Mostik says it has cracked a strange new way to get different AI models working together: skip the words entirely. Instead of passing prompts or text back and forth, Mostik's system lets models exchange raw internal signals directly, a move that could reshape how companies stitch together specialized AI systems, according to a new report from Wired.
Something odd is happening in a small AI lab tied to Russian mathematicians, and it could quietly upend how AI companies think about combining models. A startup called Mostik has developed what it describes as a way for AI models to talk to each other without using words at all, according to a report from Wired. Instead of the usual approach, where one model generates text and feeds it as a prompt to another, Mostik's system reportedly lets models pass information through their internal, non-verbal representations, essentially letting one model's 'thoughts' feed directly into another's.
If that sounds abstract, it's because right now, it kind of is. Wired's report is thin on technical specifics, and Mostik hasn't published a detailed paper or benchmark suite that outside researchers can independently verify. But the core idea taps into something the AI industry has been chasing for years: how do you get the strengths of multiple specialized models, one great at reasoning, another at code, another at vision, without the overhead of translating everything into and out of natural language every step of the way.
That overhead is not trivial. Every time a model has to turn its internal state into text and another model has to read and reinterpret that text, you lose information and burn compute. It's a bit like two experts trying to collaborate only by mailing typed letters back and forth instead of just sitting in the same room. Mostik's pitch, as described in Wired's coverage, is that skipping the letter-writing entirely, and letting models exchange something closer to raw signal, could make multi-model systems faster and more capable at once.
This isn't happening in a vacuum. The broader AI industry has been circling this exact problem for a while, through techniques like mixture-of-experts architectures and model merging, where weights or components from different models get combined mathematically rather than through conversation. Google and Meta have both published research on merging specialized models into single systems, and OpenAI has leaned heavily into routing user queries across multiple internal models depending on task complexity, a strategy detailed in our earlier coverage of OpenAI's model routing shift. If Mostik's non-verbal communication approach genuinely works and scales, it would sit in that same lineage, but pushing further toward direct, low-level model-to-model exchange rather than architectural tricks applied during training.
The timing matters too. Inference costs have become one of the biggest headaches for AI labs and their enterprise customers, a theme that's come up repeatedly as companies try to deploy increasingly large models in production, something we've tracked closely in our reporting on rising AI inference costs. Any technique that promises to cut the computational tax of getting multiple models to cooperate is going to get attention fast, especially from cash-strapped startups trying to compete with the deep pockets of Microsoft-backed OpenAI or Amazon's AWS-hosted model ecosystem.
Still, skepticism is warranted here. Wired's own report acknowledges the story is light on hard details, and there's no independent verification yet of how well Mostik's method performs against established baselines, or whether it's a genuinely novel technique versus a repackaging of existing ideas like knowledge distillation or embedding-space communication that researchers have experimented with for years. Extraordinary claims about AI breakthroughs are common right now, and plenty of startups have overstated results only to quietly walk them back once bigger labs or academic groups take a closer look.
What happens next will tell us a lot. If Mostik releases benchmarks, open-sources part of its system, or attracts serious funding from investors who've seen the technical details firsthand, that's a signal this is more than a clever pitch. If the story fades without follow-up, it's worth remembering how many AI startups have promised radical shortcuts to model performance only for the claims to not hold up under scrutiny. Either way, the underlying problem Mostik says it's solving, getting AI systems to cooperate without the friction of natural language, is real and unsolved, and whoever cracks it convincingly is going to have labs like Google and OpenAI paying very close attention.
For now, Mostik's non-verbal model communication trick is more intriguing headline than proven technology, but it lands squarely in one of the AI industry's most pressing unsolved problems: how to get multiple specialized models cooperating efficiently without burning compute on constant text translation. Readers should treat the claims with healthy curiosity rather than certainty until independent benchmarks surface, but keep an eye on this space. If a small startup out of Russia genuinely found a shortcut that the likes of Google, Meta, and OpenAI haven't cracked at scale, it wouldn't be the first time a scrappy outsider forced the giants to rethink their playbook.
