Implementing a fast and furious spellchecker when the dictionary has 10 billion terms

Question

I am considering to use SymSpell, kbtree, or some hybrid model between both for very fast fuzzy lookup, which means that given a binary sequence of length 50 (e.g. 101