When it comes to tokenization, yeah, it does differentiate. But its training set is certain to have many examples of different capitalizations and misspellings, so it can understand them, too. Which is usually what you want...
GPT-4 really takes this to eleven. For example, it can correctly parse and process this request: "Tll m shrt str bt ncrns; 10 sntncs. Wrt t l vwls prprl."
just run the 13b model 4bit quantized locally, it's already better than the 7b-8bit and you can turn down the temperature to 0 to get repeatable results.
> How long is a piece of string, when the string is 4Kim long?
...or even just Km. I can't, as it's being slashdotted