Same model, same question - different answer if loaded directly vs as openai endpoint

[hugalafutro]

Hi,
the model I use is L3-8B-Stheno-v3.2-abliterated.i1-Q4_K_M.gguf. The question I use is How many R's are there in the word strawberry (I know why most models fail it and that it's not really a good test of LLM, but disregard that)

If I load the model in the app directly, it counts 1 R most of time, sometimes 2, never seen it count 3.
If I load the model in koboldcpp, and connect gpt4all to the openai endpoint of koboldcpp, without fail it will count 3 R's and often also spell it out and explain it. (By without fail I mean I tried like 5 times with each method, which you might argue is not enough of a statistical sample, but we ain't doing statistics here, this is simple arithmetic/spelling most pre-schoolers can do.)

Why? The backend shouldn't matter, no? OpenAI endpoint is just an access to the model. So why the difference?