For a reasonable amount of literals, I suggest to save the resources in the DB or in a .RESX file in the server. When the user logs in or you detect the language that will be used, the literals are requested by the application and saved either in a collection of your translation module or in the LocalStorage of the browser (this could be a good approach for large data).
Then this module could have some methods to retrieve the messages, probably passing a key.
Using this solution you could inject this module in the viewmodels that need to show translated literals and acces them through the view:
<p data-bind="text: resourceManager.get('M01')"></a>
For large applications that would require huge localization data to be transfered, maybe some kind of modularity could be applied and only load the resources really needed for each module/section.
I don't think making recurrent requests to the server to get the translated literals is a good practise. SPA's should provide a good user experience and loading the translated literals from the server could be a blocking issue. Text is not like an image, you can render a page without all the images loaded, imagine rendering a page without the text :o
Anyway, I think the best solution would be to keep the server as repository and create a custom JS module that takes care to get data in one or multiple loads and is able to store it somewhere in the client.