Paul-André Melliès has quite an interesting paper on this topic:
...but phrased in the more general terms of T-algebras of a pseudomonad. The idea is that a pseudomonad on a 2-category (especially Cat), let you put algebraic structures on categories the same way monads let you put them on objects of a category, like sets. This is motivated by the need to put PROPs, PROBs, PROs, Lawvere theories, etc. all under one roof.
He begins by talking about how a T-algebra homomorphism (a monoidal functor in the case where the T-algebras are monoidal categories) j : A -> B induces a forgetful functor U_j from Models(B,C) to Models(A,C) in the way you mentioned. Looking for left adjoint to U_j amounts to looking for a way to push some functor backwards along j in a suitably natural way. As Tom already mentioned, this is the left Kan extension. This process is functorial, and usually written Lan_j : [A,C] -> [B,C]. Furthermore, Lan_j -| U_j.
But if we were done there, all PROPs would have free algebras, which we know is not true in general (cf. bialgebras). The hard part is proving the Lan_j is a T-algebraic left Kan-extension. In the case of Lawvere theories, this is easy, because the product structure guarantees all natural transformations of cartesian functors are cartesian, but in the monoidal case, this stuff all needs to be checked.
This is where the story starts to get more complicated. It seems quite tricky to come up with suitably weak conditions under which Lan_j is T-algebraic. Mellies phrases these in terms of distributers (aka profunctors, modules, depending on who you ask and what country you are in :-P). If functors are like functions, this are a bit like relations. The nice thing about them is they always come in adjoint pairs f_* and f^* for any functor f.
So, thm 1 in the paper is (roughly) this. If j and j^* are T-algebraic in the suitable 2-categories, C is (T-algebraically) complete and co-complete, and for any model f : A -> C, f_* o j^* factors through the up-star of the Yoneda embedding y : C -> Psh(C), then U_j has a left adjoint computed as Lan_j that is indeed the free functor.
This is quite heavy-duty (pro-arrow equipment, ends, etc.), but it seems to get the job done. It would be nice to see more concrete/specific examples of this.