russian mathematicians teach ai models to communicate without words
mostik's new approach enables ai models to interact via mathematical values, boosting efficiency and performance. this could help open-weight models compete with proprietary ones.
mostik, a russian startup founded by mathematicians, has created a technique for ai models to interact without producing text output. by using mathematical values from the models' weights, they can combine capabilities and improve performance. for example, they bridged a 753 billion parameter glm model with a 4 billion parameter qwen model, achieving halfway performance between the two at one-twentieth the cost. this approach could increase the value of open-weight models and allow them to better compete with proprietary models from frontier labs. mostik's method may also enable pairing frontier models with domain-specific models, leading to more specialized models being trained.