This paper presents an industry-driven solution for extreme multi-label classification with a massive label collection. The proposed approach incorporates a large number of binary classification models with label pre-filtering and employs methods and technologies shown to be applicable in industrial scenarios where high-end computational hardware is limited. The system is evaluated on an Estonian newspaper article dataset which contains almost 2000 unique labels and has shown to perform over 80 times faster than applying all the binary models of the entire label set without negative impact on prediction scores.