The latter one. The network is trained in full precision (this is required for the gradient calculation), but the weights are nudged towards the quantized values.