I took a number theory class at University of Cincinnati that was taught using a modified Moore method. The class size was pretty small, and credit-wise it was an "upper level mathematics elective" (i.e. not a core class), so our professor had some room to be experimental.
Like with the Moore method, the students presented all course material. The instructor sat in back, asked questions every so often if the student was not clear enough, helped us if we got impossibly stuck, and participated in the discussion when he had something insightful to say. We rotated teaching between three groups of two or three people. Our presentations were allowed to last multiple class sessions.
The class differed from the classic Moore method by loosening the requirement that students prove all theorems themselves, to address the time/constraints issue. We did have a book, but it was a computer science oriented text that left large holes in the proofs (or omitted them completely). We had to complete proofs at least well enough to field questions from each other and the professor, which often forced us to create our own lemmas. If anyone couldn't prove something at the board, the whole class had to finish it for homework. We were also encouraged to come up with our own examples, and add in whatever cool applications, algorithms, or observations we could find. Among other things, I ended up proving Chebyshev's theorem (the one leading up to the prime number theorem) after a whole bunch of lemmas, at a depth which I never would have otherwise encountered.
It turned out that most students learned the material from their own sections very deeply, but did not understand others' presentations as well as their own. I guess that's inevitable. Another complication was that some students were better at communicating their ideas than others. I bet that this method would work particularly well with a group of motivated students that had some previous practice working together. In any case, I would say I learned more in this class than other traditional classes at around the same level, so I support this adaptation of the method. It could use more development, but at the very least, I don't think it turned out worse than standard instruction.