If the overwhelming majority of words seen don't match, you would win by rejecting non-matches quickly. If instead it almost always matches, you will need to look at all the bytes anyway to be certain of a match.
If using a bloom filter with a small set, it's possible to obtain a low probability of false positives by using just one hash function and a small number of buckets. At that point you've effectively got a hashset like one of the solutions described in the blog post.
A separate hash table for what gets past would check for false positives, using the final hash value.