We discuss the implementation details and the numerical performance of the recently introduced nonconforming Trefftz virtual element method [37] for the 2D Helmholtz problem. In particular, we present a strategy to significantly reduce the ill-conditioning of the original method; such a recipe is based on an automatic filtering of the basis functions edge by edge, and therefore allows for a notable reduction of the number of degrees of freedom. A widespread set of numerical experiments, including an application to acoustic scattering, the h-, p-, and hp-versions of the method, is presented. Moreover, a comparison with other Trefftz-based methods for the Helmholtz problem shows that this novel approach results in robust and effective performance.combines the VE technology with the Trefftz setting in a nonconforming fashion (à la Crouzeix-Raviart) following the pioneering works on nonconforming VEM for elliptic problems [5,14] and their extension to other problems [3, 13-15, 25, 34, 36, 46]. This nonconforming Trefftz-VEM, which can be regarded as a generalization of the nonconforming harmonic VEM [36], is "morally" comparable to many other Trefftz methods for the Helmholtz equation such as the ultra weak variational formulation [16], the wave based method [22], discontinuous methods based on Lagrange multipliers [24] and on least square formulation [39], the plane wave discontinuous Galerkin method (PWDG) [27], and the variational theory of complex rays [41]; see [31] for an overview of such methods.It has to be mentioned that all of the above Trefftz methods are based on fully discontinuous approximation spaces. A peculiarity of the nonconforming Trefftz-VEM is that a "weak" notion (that is, via proper edge L 2 projections) of traces over the skeleton of the polytopal grid is, differently from discontinuous methods, available.The aim of the present paper is to continue the work begun in [37], where the nonconforming Trefftz-VEM was firstly introduced, an abstract error analysis was carried out, and h-version error estimates were derived. As already mentioned in [37], the original version of the method does not result in good numerical performance, mainly because of the strong ill-conditioning of the local plane wave basis functions.The scope of this contribution is manifold. After introducing the model problem and extending the original nonconforming Trefftz-VEM in Section 2, we discuss the implementation details of the method in Section 3. We will consider here a more general Helmholtz boundary value problem than originally done in [37], which will be reflected in the definition of the nonconforming Trefftz-VE spaces. Then, numerical results are presented in Section 4, in order to clarify that, rebus sic stantibus, the method severely suffers of ill-conditioning. A numerical recipe based on an edgewise orthonormalization procedure to mitigate this strong ill-conditioning is presented in Section 5. Additionally to the fact that the condition number of the resulting global matrix significantly improves, the number...