DyNet is a neural network library designed for networks with dynamic structures that change per training instance. It is written in C++ with Python bindings and optimized for both CPU and GPU execution. The library has been widely used in natural language processing applications including syntactic parsing, machine translation, and morphological inflection.