We present an extended ultraviolet-blue (850-4700Å) library of theoretical stellar spectral energy distributions (SEDs) computed at high resolution, λ/∆λ= 50 000. The Uvblue grid, as we named the library, is based on LTE calculations carried out with Atlas9 and Synthe codes developed by R. L. Kurucz and consists of nearly 1800 entries that cover a large volume of the parameter space. It spans a range in T eff from 3000 to 50 000 K, the surface gravity ranges from log g= 0.0 to 5.0 with ∆ log g= 0.5 dex, while seven chemical compositions are considered: [M/H]= −2.0, −1.5, −1.0, −0.5, +0.0, +0.3 and +0.5 dex. For its coverage across the H-R diagram, this library is the most comprehensive one ever computed at high resolution in the short-wavelength spectral range, and useful application can be foreseen both for the study of single stars and in population synthesis models of galaxies and other stellar systems.We briefly discuss some relevant issues for a safe application of the theoretical output to ultraviolet observations, and a comparison of our LTE models with the NLTE ones from the Tlusty code is also carried out. NLTE spectra are found, in average, to be slightly "redder" compared to the LTE ones for the same value of T eff , while a larger difference could be detected for weak lines, that are nearly wiped out by the enhanced core emission component in case of NLTE atmospheres. These effects seem to magnify at low metallicity (typically [M/H] −1).A match with a working sample of 111 stars from the IUE atlas, with available atmosphere parameters from the literature, shows that Uvblue models provide an accurate description of the main mid-and low-resolution spectral features for stars along the whole sequence from the B to ∼G5 type. The comparison sensibly degrades for later spectral types, with supergiant stars that are in general more poorly reproduced than dwarfs. As a possible explanation of this overall trend, we could partly invoke the uncertainty in the input atmosphere parameters to compute the theoretical spectra. In addition, one should also consider the important contamination of the IUE stellar sample, where the presence of binary and variable stars certainly works in the sense of artificially worsening the match between theory and observations.