A prototype predictive controller, based on the theory of optimal stochastic control, was developed and installed in a non-residential building in Delemont (Switzerland). This building is very well insulated (U = 0.65 W/m(2)K for windows) and, moreover, equipped with an active floor heating and cooling system. Solar and free gains supply more than 50% of its heating energy during winter. The controller determines, with the aid of the predicted free gains, the heat to supply for the next hour in order to optimise comfort and minimise energy consumption over a period of 24 h. The performance of this controller is compared with that of an advanced external temperature controller installed in the same building. A detailed energy balance of the building is presented for both controller types. The maximal energy savings observed in favour of the predictive controller amount to 24% during the cold season (October through April) and even reach 31% during the hot season (May through September). The thermal comfort was evaluated through questionnaires filled in by the occupants, but also thanks to an analysis based on temperature monitoring in the building. The analysis results show that comfort was maintained by the predictive controller.