jats:titleABSTRACT</jats:title> jats:pWe present constraints on primordial B modes from large-angular scale cosmic microwave background polarization anisotropies measured with the Planck satellite. To remove Galactic polarized foregrounds, we use a Bayesian parametric component separation method, modelling synchrotron radiation as a power law and thermal dust emission as a modified blackbody. This method propagates uncertainties from the foreground cleaning into the noise covariance matrices of the maps. We construct two likelihoods: (i) a semi-analytical cross-spectrum-based likelihood-approximation scheme (momento), and (ii) an exact polarization-only pixel-based likelihood (pixLike). Since momento is based on cross-spectra, it is statistically less powerful than pixLike, but is less sensitive to systematic errors correlated across frequencies. Both likelihoods give a tensor-to-scalar ratio, r, that is consistent with zero from low-multipole (2 ≤ ℓ &lt; 30) Planckpolarization data. From full-mission maps, we obtain r0.05 &lt; 0.274 at 95 per cent confidence at a pivot scale of k* ≡ 0.05 Mpc−1, using pixLike. momento gives a qualitatively similar but weaker 95 per cent confidence limit of r0.05 &lt; 0.408.</jats:p>