A little-known Russian startup called Mostik is reportedly working on a strange new way to get different AI models to cooperate, not by having them exchange text or code, but by letting them pass information to each other in a form that skips human language entirely. The approach, first reported by Wired, could reshape how companies combine specialized models instead of building ever-bigger single systems.
Every major AI lab right now is wrestling with the same problem: how do you get multiple models, each trained for a different job, to actually work together well? The usual answer has been to stitch them together with text. One model writes out its reasoning, another model reads it, and everyone hopes nothing gets lost in translation. A startup called Mostik thinks that's the wrong way to do it entirely, and according to a report from Wired, the company has built a way for models to pass information to each other without going through language at all. Instead of one model explaining itself in English or code, the idea is that models can share something closer to raw internal representations, the kind of dense numerical signals models use internally, cutting out the lossy step of translating everything into words first. It's a genuinely unusual pitch, and it comes from a team the report describes as mathematicians rather than typical Silicon Valley founders, which tracks with how technical and almost academic the approach sounds. This isn't happening in a vacuum. Every big player in AI has been pushing toward some version of multi-model collaboration over the past two years. OpenAI has leaned into what's often called mixture-of-experts architecture, where a system routes different tasks to different specialized sub-models under one roof. Google has done similar work with its Gemini family, and Meta has published extensively on model merging techniques that blend the weights of separately trained networks into one. What sets Mostik's approach apart, at least based on what's been reported so far, is that it's not trying to merge models into a single blob or route tasks through a central manager. It's trying to let fully separate models talk to each other directly, in something like their own native format, while staying independent systems. If that works reliably, it would be a meaningfully different path from the merge-everything-into-one-model strategy that's dominated the last few years of AI development. There's an obvious reason companies want this to work: training a frontier model from scratch is absurdly expensive, and even fine-tuning existing ones takes real compute and real money. If a smaller lab could instead take a model good at, say, medical text and pair it with one that's strong at math reasoning, and have them cooperate without retraining either one, that changes the economics of building AI products. Smaller shops and startups without the budgets of Microsoft or Amazon could suddenly punch above their weight by combining off-the-shelf models instead of building their own from zero. That said, there's a lot we still don't know. The original report is thin on specifics, no word yet on funding, headcount, which models Mostik has tested this with, or whether any enterprise customers are already using it. Reasonable skepticism applies here too. AI research has seen plenty of promising-sounding interoperability tricks that worked beautifully on a narrow benchmark and then hit a wall the moment someone tried to scale them across model families with different architectures and training data. Getting two models trained by completely different teams, on different data, using different tokenizers, to meaningfully exchange internal representations is a genuinely hard technical problem, and it's one that bigger, better-funded labs have been chipping away at without a clean solution. So the real test for Mostik will be whether this holds up outside a lab demo. Does it work across model families from different vendors, not just variants trained in-house? Does it scale past toy examples? And can a small startup actually commercialize something this technical before a bigger player builds a competing version with more resources behind it? Those are the questions worth watching as more details, hopefully, come out.
For now, Mostik's idea is more intriguing headline than proven product, a single Wired mention with big implications if it pans out and plenty of open questions if it doesn't. But it lands at a moment when the entire industry is hunting for cheaper, more efficient ways to combine AI capabilities instead of just building bigger models, so it's worth keeping an eye on whether this Russian startup's math-heavy approach to model-to-model communication actually makes it out of the lab and into real products.
