The models that power LinkedIn’s personalized ranking (e.g. for job postings, feed updates) may use billions to trillions of features. To train these models efficiently on commodity hardware, LinkedIn developed the Generalized Deep Mixed Model (GDMix). GDMix is a framework implemented in Tensorflow, Scipy, and Spark to train nonlinear fixed effect and random effect models, which often power search and recommendation systems. It splits a larger model into a global fixed effect (that sets the global trend) and a large number of small random effects (which account for individuality), which it solves individually. Supported models include logistic regression and DeText models; however, users can also work with custom fixed effect models that are not natively supported by GDMix.