Wouldn't that almost certainly lead to measurable loss of capabilities?

If the model was quantized/distilled correctly, not for a large swath of use cases/problem domain. For anything where loss was not measured during distillation, very likely.