You want to test if the response in a voxel ('r') is reasonably modeled as being like
r(p) = a + b*(p-pbar) + c*(p-pbar)*(p-pbar)
where p = your parameter (pbar = mean p), and a,b,c = to be estimated -- is this correct?
_AM2 isn't set up for this model, with 3 unknowns per voxel -- it is for 2 unknowns per voxel.
One alternative would be to use -stim_times_IM to get the response estimate r for EACH stimulus (save them via -cbucket), then do the quadratic fitting/testing separately treating those response estimates as the 'data' to be modeled. That could be done using 3dDeconvolve again, using -stim_file to set up 3 regressors: constant, linear, quadratic in 'p' (and using '-polort -1' so there is no baseline model!).
The clarity level of the above paragraph depends on the depth to which you understand how 3dDeconvolve works, I'm afraid. Not knowing your status in the Jedi AFNI Master training, I'm not sure how much more detail to provide.
The second alternative would be to convince me to rewrite the code to have an _AM3 (_AM4 ...?) option, but that might be difficult. That would take a couple days, and would be dull.