A team of mathematicians working at an artificial intelligence startup has developed a novel method for AI models to communicate directly with one another. The startup, named Mostik after the Russian word for bridge, designed an approach that links distinct neural networks by using the raw mathematical values embedded within their internal weights. In machine learning, weights are the underlying numerical parameters that dictate how an input prompt is transformed into a generated response. By establishing a direct connection between these values, the capability of a massive frontier model can be shared with a smaller companion model, upgrading its performance rapidly and at a dramatically reduced computational cost.
ARC-AGI 3 Success and Benchmark Performance
Using this weight-bridging technique, Mostik developed an AI system that reached the top of the ARC-AGI 3 benchmark leaderboard, an evaluation widely regarded as one of the toughest tests for machine intelligence. While the startup has kept the full architectural specifics confidential due to the ongoing contest, they built a public demonstration linking two open-weight models developed in China. The team connected the largest iteration of GLM-5.2, which contains 753 billion parameters, with a 4-billion-parameter version of Qwen-3.5 designed to execute locally on consumer smartphones. The resulting hybrid system achieved performance exactly midway between the small and large models, while operating at merely one-twentieth the computational cost of the full GLM model.
Bypassing Text Generation Bottlenecks
It is widely recognized across machine learning that combining outputs from multiple models, a practice known as ensembling, produces superior results compared to running a single model in isolation. Mostik Chief Executive Officer Sasha Malysheva compared the phenomenon to a classic mathematical demonstration regarding collective estimation. In statistics, combining and averaging independent guesses from a crowd of non-experts often yields a more accurate measurement of an object's weight than relying on a single expert opinion. Combining output from several AI models functions similarly. However, traditional ensemble techniques rely on generating text from one model and feeding it as a prompt into another, an operation that consumes significant time, memory, and energy. Mostik eliminated this text generation bottleneck by enabling silent weight-level mathematical communication. If widely adopted, this strategy could substantially increase the utility of open-weight systems competing against proprietary models from companies like Anthropic and OpenAI.
Shift Toward Specialized Domain Architectures
Sasha Malysheva believes that coupling diverse, smaller models will prove to be a more viable path forward for artificial intelligence than relying strictly on raw scaling. Rather than building massive monolithic models trained on vast internet datasets, future improvements may rely on combining general-purpose networks with specialized systems tailored for fields such as biology or physics. Vladimir Arustamian, tech lead at the AI software company Lovable, noted that Mostik's team managed to deliver a working weight-bridging pipeline in a matter of months, delivering progress that observers expected would take years of research.
Efficiency Gains and Brain Mechanism Comparisons
Karl Tuyls, a former computer scientist at Google DeepMind, highlighted that Mostik's framework allows developers to approximate the capabilities of giant models without running the entire inference loop through them. By running a compact model alongside a bridged network, operations achieve major performance uplifts while staying resource-efficient. Meanwhile, Stanislav Smirnov, a professor at the University of Geneva and recipient of the 2010 Fields Medal who serves as Mostik's chief scientist, emphasized the theoretical significance of the work. Smirnov noted that finding a common mathematical framework between disparate neural networks is extremely challenging because formal mathematical languages for model interaction do not yet exist. Bridging the gap empirically provides a practical solution and could offer new insights into how artificial models function compared to human cognition. Deeper mathematical study may eventually reveal shared patterns in how artificial neural networks and the human brain process complex reasoning tasks.
Overcoming Doubts and Academic Skepticism
Sasha Malysheva's interest in mathematics began in her youth after her older brother suggested she lacked the skill to solve Math Olympiad problems. Motivated to prove him wrong, she mastered advanced mathematics and secured admission to an elite educational institution in St. Petersburg. When she later proposed the concept of creating a direct mathematical bridge between AI model weights, several industry peers expressed skepticism, suggesting the task would be too complex to execute. By turning that skepticism into drive, Malysheva and her team successfully demonstrated that direct weight communication can make frontier-grade intelligence far more accessible and cost-effective across the industry.
