Hi Nikolai, I have seen the same thing with the different dispersion data sets in the relax test suite. The numerical model from the Maple expansion appears to be rock solid. Well, at least your implementation which is now in relax. It's only disadvantage is computation time - though much faster than the standard numeric solutions (also in relax), it is still much slower than the analytic models. But if you take the CR72 model result as a starting point for optimisation - as is performed in the auto-analysis of relax - then that disadvantage pretty much disappears. One exception is if the CR72 model result is no good, then it'll take a lot more time to reach the real minimum. If it holds for all timescales and populations, then Andy's model should theoretically find the same result as the numerical result. It would take a comprehensive study over the entire parameter space to demonstrate that this holds everywhere, but this is unfortunately missing from the paper. But as it is an analytic model rather than numeric model, then it will have the advantage of speed. So if it were to be added to relax and someone is interested, it should be very easy to perform a fair comparison of speed and accuracy between all the different models. Speed is also very much implementation dependent and small rearrangements of the order in which calculations are preformed can make an order of magnitude difference. So I don't understand where the factor of 130 mentioned in the abstract comes from. Anyway, if Andy's model finds the same parameter values and errors as the numerical models (excluding the off-resonance effects) then it could theoretically replace all 2-site analytic and numeric CPMG dispersion models. Regards, Edward