I'll try to explain using an example from my experience.

The plane's parameters is the normal and the distance to the plane from a fixed reference frame. The model's (plane) coefficients, in this case, are the normal vector (x,y,z) and the offset.

The model's coefficients are estimated using a least-squares method inside RANSAC to find the best-fit model. Once those coefficients are found; an optimization is performed over the estimated coefficients to minimize or reduce the mean-squared-error with respect the all of the 3D points.

The optimization a bit longer and can fail if you have a large percentage of outliers.

This method is optional due to the fact that if you know you have no outliers and no noise, then the least-squares solution will be the same as the optimized (refined) solution. (Saves computational time)

However, if you know that your points are affected by noise and you had to remove outliers, then optimizing the model's coefficients will give a better solution with respect to the least-squares solution. (More time, but better solution)

Its a trade-off between time and quality of the solution.