System identification for large scale continuous time nonlinear state space models, as common in mechatronics applications, remains a challenging task. Existing system identification methods can be deficient for complex nonlinear problems and computationally do not scale well on large problems, the focus of this work is to alleviate these issues through the use of Gaussian mixtures. This allows for nonlinear and non-Gaussian state estimation, as possessed by particle based approaches for example, whilst having the advantage of remaining computationally efficient for higher order systems. This talk will present some of the limitations in existing system identification methods, the approach being taken to overcome these difficulties, and some preliminary results.