The paper is devoted to model based testing against non-deterministic specifications. Such test derivation strategies are well developed, for example against non-deterministic Finite State Machines, however the length of the corresponding test suite can be exponential w.r.t. the number of specification states. We therefore discuss how a test suite can be minimized or reduced when certain level of guarantee concerning its fault coverage is still preserved. The main idea behind the approach is to augment the specification by assigning probabilities for the non-deterministic transitions and later on evaluate the probability of each test sequence to detect the relevant faulty implementation. Given a probability P which is user-defined, we propose an approach for minimizing a given exhaustive test suite TS such that, it stays exhaustive with the probability no less than P.