Your argument is correct. In fact, we can axiomatize any finite model, meaning that if ${\mathcal M}=(M,\dots)$ is a model (in a language of arbitrary size) then there is a set of sentences $\Sigma$ such that any model of $\Sigma$ is isomorphic to ${\mathcal M}$:
Say $M=\{a_1,\dots,a_n\}$.
First you say that the universe has size $|M|=n$, with a sentence $\tau$ such as "there are $x_1,\dots,x_n$ which are pairwise different and such that each $y$ is equal to one of them."
For each relational symbol $R$, consider the formula
"there are $x_1,\dots,x_n$ pairwise different and such that $\bigwedge_{R^{\mathcal M}(a_{i_1},\dots,a_{i_k})}R(x_{i_1},\dots,x_{i_k})$ and such that $\bigwedge_{R^{\mathcal M}(a_{i_1},\dots,a_{i_k})\mbox{ fails}}\lnot R(x_{i_1},\dots,x_{i_k})$",
where the first big conjunction ranges over all tuples of elements of $M$ that are in the interpretation of $R$, and the second runs over all tuples that are not in that interpretation. Given any model of this formula of size $n$, there is a bijection between this model and $M$ such that the interpretation of $R$ in this model is just the image under the bijection of the interpretation of $R$ in ${\mathcal M}$. Let $\phi_R$ be the formula such that the sentence we just wrote is $\exists \vec x\phi_R$.
Similarly, there is a sentence $\exists\vec x\psi_f$ describing completely each function $f^{\mathcal M}$, in the same sense as the sentence above completely describes $R^{\mathcal M}$.
$\Sigma$ consists of $\tau$, and the following formulas:
For any finitely many constant symbols $c_1,\dots,c_j$, any finitely many relational symbols $R_1,\dots,R_s$, and any finitely many function symbols $f_1,\dots,f_k$, say that $c_l^{\mathcal M}=a_{i_l}$ for each $l$, the sentence
"$\exists \vec x(\bigwedge_{a=1}^s\phi_{R_a}\land\bigwedge_{b=1}^k\psi_{f_b}\land \bigwedge_{l=1}^j c_l=x_{i_l})$".
Note that if the language is finite, $\Sigma$ is finite as well. But the theory just described uniquely characterizes ${\mathcal M}$ even if the language is infinite.
This result is usually presented as a consequence of Beth’s definability theorem or of Svenonius’s theorem, see for example W. Hodges, Model theory. Cambridge University Press, (1993), Chapter 10.
(I guess this says that there is no useful compactness result for finite models, since the only version that would hold would simply give us a model that we already knew we had.)
