LLM generations of that length tend to go off the rails - I think generating it in chunks where you can try and guide the model back onto the rails it probably a more sane technique.

There are several open source llms to lean on - but for long generations you’ll need a lot of memory if you’re running it locally.