With the advancement of computer graphics in the recent years, an increasing number of pictures, video and 3D content is generated by synthesis processing rather than acquired with capture devices such as cameras or scanners. Several techniques have been developed for compression of discrete (i.e. piece-wise planar) 3D models, in the form of 3D polygonal meshes. However, no important attempt has been made to compress the smooth surfaces of artificially generated 3D models, that are most often represented as parametric surfaces, of which Non-Uniform Rational BSpline (NURBS) is a popular form. This paper presents a method for compressing NURBS 3D models with a small and controllable loss. The scheme uses a differential pulse coded modulation (DPCM) coder with different predictors for knot values and control points, coupled with a uniform scalar quantizer, followed by a bitplane arithmetic entropy coder. The multiplicity of knots is preserved by the use of a multiplicity map. The rate-di...