Having read the original research on word embeddings published by Google and others, I'm sad to say that it made no mention of how to handle such data. While you could hypothetically just run a model over the raw, unedited text, depending on your desired end-use (and the dimensionality of your data), preprocessing might be highly beneficial.
If I was trying to establish a word-embedding model and didn't have enough raw data containing numbers to allow the model to figure out differences between different numbers and their uses, I'd likely use regular expressions to help simplify the inputs. While you could replace all numbers (of any length) into a uniform label (i.e. 'NUM' as @HFulcher suggested), depending on your application you may be losing data by doing this. You may want to differentiate between prices, fractions/percentages, ordinal numbers, dates, etc. This can be done by varying the labels that you replace the raw numbers with ('PRICE', 'DATE', etc). Hypothetically you could also swap between numerals and spelled-out numbers if they fit a certain category (like falling under a certain ceiling or being ordinal).