Mostik Links GLM and Qwen Models to Cut AI Costs

The startup Mostik has introduced a novel ensembling technique that allows different artificial intelligence models to interact directly using the mathematical values in their weights, bypassing the need to generate text outputs. To demonstrate the approach, the company built a bridge between two Chinese open-weight models: the largest version of GLM-5.2, which features 753 billion parameters, and a mobile-friendly 4-billion-parameter version of Qwen-3.5. The resulting hybrid system operates at just one-twentieth of the cost of the full GLM-5.2 model, while its performance lands exactly halfway between the two.

Mostik has also applied this method to a model that reached the top of the ARC-AGI 3 competition leaderboard. Led by CEO Sasha Malysheva and chief scientist Stanislav Smirnov, a 2010 Fields Medalist, the team aims to challenge the industry's reliance on scaling monolithic models. Malysheva noted that combining model outputs typically requires feeding text from one to another, which is slow and expensive. By enabling direct weight-based communication, Mostik's technique allows open-weight models to compete more effectively against proprietary systems from frontier labs like OpenAI and Anthropic.

For machine learning practitioners, this development offers a highly efficient way to boost the intelligence of smaller, domain-specific models by pairing them with frontier models. Instead of running a massive, expensive model for an entire workflow, developers can run a smaller model alongside it to achieve near-large-model quality at a fraction of the computing cost. According to industry experts, this approach could accelerate the training of specialized models in fields like biology and physics, making high-performance AI deployment far more accessible and cost-effective.