In a different session, we repeated the experiment, randomized block
design, but now we run A1B1 against A1B2. No fixation.
By 'no fixation' do you mean that there were no fixation periods in your experiment design with two stimuli A1B1 and A1B2? You would have to add some fixation periods (30% - 50% of the total scanning time) in the design so that regression analysis can estimate the baseline and any drifting effect with reasonable confidence.
Have you tried simple regression with 3dDeconvolve by modeling the impulse response function with some standardized format such as gamma variates (via AFNI program 'waver') ? With block design, simple regression may generate more meaningful and better results than deconvolution.
Gang