Designing read-back subsystems in magnetic recording requires precise knowledge about the signal picked up by the reading head. As areal densities in longitudinal magnetic recording increase, the read-back signal becomes more corrupted by intersymbol interference, media noise, and intertrack interference. Due to the spatial distribution of transitions on a disk surface and the nonlinear character of bit interactions, two dimensional media plane models are widely used to model the write process. If a two dimensional (or three dimensional) read- head model is utilized, intertrack interferences are also observed. Micromagnetic media modeling, coupled with appropriate read-head models, have been successfully used to model the 'raw' magnetic recording channel. However, due to its high computational complexity, micromagnetic modeling is an impractical tool in statistical signal analyses such as error rate studies where thousands of transitions need to be created. We propose a much simpler, yet realistic, two dimensional write process model. We call it the triangle zig-zag transition (TZ- ZT) model since the transition boundary is modeled by lateral sides of isosceles triangles of alternating orientations truncated on a common basis line across the track width. Formulas are presented that relate the parameters of the model, the probability density function of triangle heights and the constant vertex angle, to the magnetization transition profile of an isolated transition and to the cross track correlation width, respectively. Although stochastic zig-zag models have been proposed in the past, our model has the advantage that it is stable across the track, that is, it is not an independent increment process and it therefore doesn't exhibit a cross track drift. Compared to micromagnetic modeling, the TZ-ZT model offers computational savings of 4 orders of magnitude, while transition shapes and media noise are modeled with comparable accuracy, as our results show. For these reasons, the TZ-ZT model, combined with an appropriate head-sensitivity function, is an attractive 'raw channel' model for applications such as statistical performance analyses where large numbers of bits are needed.