Are you using a recent version of 3dDeconvolve? I though I had fixed this problem with the constant term estimate, by removing the mean from all the baseline components except the 1(t) term.
Of course, if the stimulus regressors have a large mean, that could cause a problem. Normally, I would think the stimulus regressors would be based at zero (not necessarily with a zero mean), but this should not cause a problem unless there is very little 'baseline' time -- that is, the subject is basically in some stimulus condition all the time.