Hi Vikas,
This is quite clearly outlined in Don's paper. You need to make sure you have something like this:
Code:inline subckt mynmos d g s b
parameters w=1u l=1u ad as pd ps ...
mynmos (d g s b) mynmosMod w=w l=l ad=ad ...
// parameterised model
model mynmosMod bsim3v3 type=n param1=value1 param2=value3 ...
ends mynmos
The idea is that you presumably have some parameterised model, with a statistics block describing the std deviation and distribution type of the parameters referred to in the model. What you then do is surround this with a subckt definition (it doesn't have to be inline; that just helps with backannotation), and then you can use the "model" mynmos wherever you like, as if it were a model. Because the real model (mynmosInt) is a parameterised model within a subckt, each instance has its own model, and so they will not be shared. If all instances are shared, you can't have mismatch - whereas if all models are unique to the instance, they can have mismatch.
Does that help?
Regards,
Andrew.