If I remove the pm.Potential from my model, it will be identical to yours with w=[0.5, 0.5]. But I need the potential term to not get samples that are inconsistent with what I know about how the data was generated. Without the potential, of the 60 observations, on average 30 will be assigned to the one component and 30 to the other, but by chance it usually deviates from this, and the deviations can be as large as 40/20. Whereas with the potential most samples are 30/30, and the largest deviations are 32/28.