The ability to execute Deep Neural Networks at the trigger level to improve online selection performance will be crucial for current and future high-energy physics experiments. Low-latency hardware solutions exist, e.g. FPGAs, but the primary constraint to the implementation is often related to the model's size, which has to be finely tuned not to exceed the available memory. We present here an approach to reduce the size of models, having under control the model performances. Promising results are shown in the classification problem of selecting proton-proton collision events in which the boosted Higgs boson decays to two $b$-quarks, and both the decay products are contained in a large and massive jet, against an overwhelming QCD background.