We present a new way of drawing a crowd of animated characters in real-time. Previous work has focused almost exclusively on how to visualize ever larger crowd scenes and the current state-of-the-art can display tens of thousands of virtual humans with ease. The associated trade-off, however, is that crowd members can do little more than play a set of scripted motion clips. It follows that designating individuals to be members of a crowd instantly limits the techniques that can be used, the behaviours that can be depicted and ultimately, the perceived realism of a scene. Our approach differs from the state-of-the-art in that we do not propose a crowd-specific technique but instead a bone-parallel, OpenCL-accelerated interpretation of the traditional character pipeline. The method does not rely on pre-processing; provides fine-grained control over the animation of a crowd (support for motion blending and varied skeletons, for example) and crowd members and user-controlled ‘hero’ characters can be handled without distinction. Display omitted ► We animate many thousands of virtual humans in real-time. ► The problem is modelled by a three-stage pipeline of OpenCL kernels. ► Kernels can run on the CPU, GPU or both. ► The approach provides far greater control than equivalent techniques.