There are bugs in both v11 and v12 which prevent correct convergence for some datasets with low counts in both the source and background files. The v11 fix is available as 11.3.2ab and the v12 fix will be available shortly.
However, it appears that the eventual best fit is still biased - this is likely a statistical rather than computational problem. The model seems to be be biased systematically below the data so a possible approach is to add another term which pushes the total number of model and data counts to be the same.
Keywords: HEAsoft, xspec