Teaching Models to Speak without Words

According to a report from WIRED, a Russian startup called Mostik—founded by CEO Sasha Malysheva alongside Fields Medalist Stanislav Smirnov—has developed a method of machine "telepathy" that enables AI models to communicate directly via their mathematical weights instead of generating slow, compute-heavy text tokens. Moving away from the brute-force strategy of building ever-larger monolithic models, Mostik creates an internal bridge between disparate systems so a massive frontier model can feed its capabilities straight into an efficient, smaller model. In a demonstration connecting a massive 753-billion-parameter GLM-5.2 with a lightweight 4-billion-parameter Qwen-3.5, the resulting hybrid delivered benchmark performance directly halfway between the two systems while slashing runtime costs to just one-twentieth of the frontier model. Having already climbed to the top ranks of the ARC-AGI 3 challenge, this weight-bridging architecture could give open-weight developers a powerful, cost-effective way to rival the closed, proprietary systems of OpenAI and Anthropic without needing massive infrastructure budgets.