To develop large-scale neural models, we need software tools that not only support efficient simulation of hundreds of thousands of neurons, but also provide a structure for constructing and organizing these large-scale models. This demonstration showcases Nengo,our neural modelling system that integrates a wide variety of neural modelling techniques. It supports user-definable single cell and synaptic models, as well as providing several common default models. More importantly, it allows for the high-level design of complex neural circuitry by allowing users to specify representational properties of neural groups and the desired transformations that should be caused by synaptic connections between them. This is a unique feature of the software. Furthermore, the synaptic weights can then either be directly solved for using the Neural Engineering Framework (NEF) methods described in Eliasmith & Anderson (2003) or the weights can be generated using user-defined local learning rules. These features can be intuitively exploited through the GUI, or extensively customized through an interpreted command line and scripting. Nengo allows users to create large-scale models by focusing on this higher-level description. This method has been successfully used to construct models of: working memory using 25-dimensional neural integrators; motor control using Kalman filters; escape and swimming control in zebrafish; the translational vestibular ocular reflex; symbolic decision-making and rule-following in the basal ganglia; and others. Uniquely, Nengo provides direct control over the efficiency and accuracy of the neural simulation by allowing for alternate neural models, switching from precise spiking models down to rate models and even to an abstract mode for directly computing the underlying transformations. Other efficiencies are implemented as well, including support for synaptic weight matricies of low rank and user-definable mathematics libraries. These aspects allow for fast simulation of large systems, making Nengo ideal for exploring interactive behaviour of complex neural models.