In video coding systems using adaptive arithmetic coding to compress texture information, the employed symbol probability models need to be retrained every time the coding process moves into an area with different texture. To avoid this inefficiency, we propose to replace the probability models used in the original coder with multiple switchable sets of probability models. We determine the model set to use in each spatial region in an optimal manner, taking into account the additional signaling overhead. Experimental results show that this approach, when applied to H.264/AVC's context-based adaptive binary arithmetic coder (CABAC), yields significant bit-rate savings, which are comparable to or higher than those obtained using alternative improvements to CABAC previously proposed in the literature.