In a perfect crystal at zero temperature the wave functions are completely delocalized.
Now, as we get more and more realistic, and consider various physical phenomena in conductors, we need to account for imperfections, interactions, non-equilibrium, etc.
Core electrons of atoms, corresponding to deep lying energy bands are pretty much in Bloch states (although strongly peaked at the ions.) Valence and conduction band electrons are rather spread - after all, they still assure the binding between the atoms, which holds the crystal together. However, various interaction/scattering processes, as well as crystal imperfections (impurities, dislocations) result in finite coherence length - the characteristic distance after which the phase of electron is randomized. This coherence length can be considered roughly as the size of electron.
The coherence length is usually much longer than the lattice constant, so that we can still use the Bloch theory (that is the energy bands theory), however compared to the sample size we can consider it very short and hence treat electrons as particles.
Nowadays there are possibilities to create semiconductor nanostructures with very large coherence lengths - of the order of microns, that is thousands or tens of thousands of lattice spacings. This allows studying various quantum effects, where conductance deviates from Drude description - such as weak localization, conductance quantization, Coulomb blockade, etc. This is often referred to as mesoscopic physics, mesoscopic structures, meaning that the structures in question are not microscopic (they are much bigger than atoms), but they cannot be treated macroscopically either, due to the essential quantum effects.
Update
At not too low temperatures we can actually estimate the conduction electron size from a simple argument: in effective mass approximation, the density matrix of free electrons is $\rho(\mathbf{k})\propto \exp\left(-\frac{\mathbf{p}^2}{2m^*k_BT}\right)$, that is the electron wave packet has momentum uncertainty $\sigma_p^2=m^*k_B T$. The position uncertainty then can be estimated as $\sigma_x^2\approx\frac{\hbar^2}{4\sigma_p^2}=\frac{\hbar^2}{4m^*k_BT}$. Electrons become more and more localized at higher temperatures.
Related question is Why do Drude/Sommerfeld models even work? - this question does not specifically focus on localization, but addresses a similar problem of why a particle description is appropriate for discussing properties of semiconductors and metals.
