The lowest possible is two  level 1 units within a level 2 unit - the classic example is each person  within a twin pair. Note in this situation you will be able to say little about a specific level 2 units( the Jones twin pair) but provided you have enough of them you will be able to says something about between and within twin variation.

You may want to look at my contribution to this exchange

https://www.researchgate.net/post/What_are_sample_size_requirements_and_considerations_in_3-level_multilevel_linear_modeling

It would also help if you can clarify what you mean by 'condition'

Thank you - I'm reading your other response too.

By condition, I mean an experimental condition which all participants undergo. To explain a bit further - I want to investigate the relative effects of 4 IVs on a single outcome. Three IVs are "between subjects" (e.g., age, gender) but the fourth IV (experimental condition: A and B) is within subjects. I understand that a multilevel model is required because of the repeated measure, rather than a typical linear multiple regression model.

Ok - condition is actually two occasions nested within an individual

So you could model it as a single level model with the later outcomes as the response and the earlier outcome is the predictor - you can then put in a dummy if a person has been treated or not - makes things a lot easier - see quant material and appendix in this

https://www.researchgate.net/publication/256802269_Evaluation_of_the_primary_schools_whiteboard_expansion_project

Article Evaluation of the primary schools whiteboard expansion project

Thank you for the very helpful answer. 

You could try it both ways and see if it makes a difference.

On the question of non-independence or autocorrelation - it is important to remember that this is a conditional assumption - that is you are assuming that Y given the X's is not autocorrelated , in effect the assumption applies to the residuals. So if you are using a single level model where the X is the past outcome you are effectively modeling a lot of dependency by inclusion of that variable.This single level approach is often used in voting studies where they condition on past voting intention and thereby focus on change.