Industrial distillation columns typically lack adequate on-line instrumentation required for implementing high performance, model-based control systems which depend on composition measurements. The typical practice is to base product quality control schemes instead on tray temperature measurements since these are usually available on-line. While this strategy has been applied with success on a wide variety of distillation columns, some special considerations may be necessary in dealing with high-purity columns. This paper presents results which address some of the key issues involved with the use of tray temperature measurements as surrogates for composition measurements, and with the development of appropriate input/output models for multivariable control of high-purity columns. For the purpose of illustration, we investigate the performance of several model predictive control schemes based on a linear model, a "high frequency" model, and a:nonlinear temperature transformation, along with that of conventional multiloop PID control. Closed-loop control performances are compared for setpoint changes as well as for changes in unmeasurable load disturbances.