The 3D Set Partitioning In Hierarchical Trees (SPIHT) for video compression is an extension of the SPIHT algorithm, which is initially introduced by A. Said and W. Pearlman for image compression. Previous works have shown that the performance of 3DSPIHT with Arithmetic Coding (AC) is comparable to H.263 and MPEG-2. Moreover, the output bit stream of 3DSPIHT is inherently embedded and scalable in rates. It is also relatively easy to make the bit stream become scalable in resolution with some minor changes. Although all these features are very attractive for certain applications that required progressive transmission or heterogeneous network, the configuration of AC can be tedious and remains as a challenging task. The changeable parameters in AC include the type (fixed or adaptive) of models, number of models, and maximum frequency to reset the models. This work presents a configuration of adaptive models in AC, which can help to improve the coding efficiency of AC for 3DSPIHT, and thus achieve better performance in terms of Peak Signal-to-Noise Ratio (PSNR). The adaptive models are used to store the probability distribution of all the symbols that appear in a system. In the proposed configuration, each type of output bits in 3DSPIHT is assigned with a separate set of adaptive models. This proposed configuration takes into account the different probability patterns which exist in each type of output bits. The maximum frequency used to reset the adaptive models is also investigated. It will not only affect the adaptation rate which directly relates to the coding efficiency of AC, but also the memory requirement. The simulation results show that the proposed configuration can improve the mean PSNR for various video test sequences in QCIF and SIF formats.