Mostik's Novel Approach Links AI Models Without Text

Mostik, a startup founded by Russian mathematicians, has developed a method for AI models to interact using the mathematical values in their weights, bypassing text output. This allows a smaller model to gain capabilities from a larger one, creating a hybrid that costs a fraction of the full model while performing halfway between the two. The approach could enhance the value of open-weight models against proprietary ones.

Mostik's hybrid pairs a 753-billion-parameter GLM-5.2 with a 4-billion-parameter Qwen-3.5, yielding performance midway between the two at roughly one-twentieth the cost. The company has reportedly topped the ARC-AGI 3 leaderboard, though it is withholding details until the competition concludes. The technique builds on ensemble principles, where combined model outputs typically outperform individual ones, but eliminates the expensive text-based handoffs between models.

Chief scientist Stanislav Smirnov, a 2010 Fields Medalist, notes that no adequate mathematical language yet exists for aligning two models' internal representations, making the bridge approach a practical interim solution. The team's work has drawn interest from researchers familiar with frontier AI, who say the method could make efficient model deployment significantly more accessible.