It is known that muscle spindles provide the majority of information about limb position, but little is known about how position sense is computed from their signals. We have developed a family of musculoskeletal models in order to determine some of the fundamental properties associated with transforming noisy spindle information into putative internal coordinate frames for position sense. A two-joint model was developed containing one biarticular and two monoarticular muscles with a total of 1000 sensors distributed among them. The sensors were assumed to function like spindle secondary afferents under fusimotor control designed to optimize their ability to encode static position in the presence of constant output noise. The optimal distribution of sensors was found to depend strongly on the coordinate frame in which position was measured (intersegmental angle, segment orientation, or end-point of the limb) and on the topology of the biarticular muscle with respect to the plane of motion. A similar analysis was performed for an anthropometric model of the human arm, using previously published counts of muscle spindles. In general, the actual distribution of spindles about the elbow and shoulder does not seem to favor any single coordinate frame for position sense. We also looked at the potential accuracy in detecting changes in joint angles based on the distribution of muscle spindles throughout the human body. The distribution of spindles about individual joints accounts well for psychophysical data showing a proximodistal descending gradient of angular resolution that partially reflects the relative importance of more proximal joints for determining the location of the end-point.