The Mw 7.2 earthquake of 21 August 2003 provides an unprecedented opportunity to evaluate the applicability of the New Zealand acceleration response spectrum attenuation model (McVerry et al., 2006) in the Fiordland region and to consider effects from 3DQ heterogeneity. This was the largest shallow earthquake in New Zealand in the past 40 yrs. It was well recorded by the GeoNet seismograph and strong-motion network. We compare the observed 5% damped response spectral accelerations (SAs) to the New Zealand model values, for both crustal reverse and subduction interface earthquakes. On average, both models underpredict the observed SAs, by approximately 0.4 in lnSA, for 0–400 km distance; however, the data show great variability. Using the South Island 3DQ model, we compute the path-averaged attenuation rate, CQ for synthetic Fiordland hypocenters to distributed sites. Compared to the standard model, the relatively high Q in South Island produces lower attenuation rates, although the low-Q crustal root of the Southern Alps produces a 10 times greater attenuation rate for certain paths. The average attenuation rate estimated from the SA data is similar to CQ for frequencies <5 Hz and constant above 5 Hz. For an acceleration model for Fiordland earthquakes, 0–60 km depth, an applicable anelastic attenuation rate is the mean CQ from the synthetic hypocenters, which is 0.006 km-1 at 2.5 Hz. For the observed SAs, the most significant regional effect is high relative SAs at some sites that may be due to postcritical Moho reflections from the Southern Alps crustal root.