On where quantisation starts showing: it is not general quality that goes first, it is instruction following on long or multi-step prompts. Short answers stay fine, which is why casual testing misses it.
Test with your longest realistic prompt, not a one-liner, or you will pick a level that fails exactly when you need it.