Joel has given an answer based on a modern view of strong compactness in terms of elementary embeddings. I'd like to complement that with an answer based on the oldest view of strong compactness of $\kappa$: The infinitary logic $L_{\kappa,\kappa}$ has the compactness property that, for any set $S$ of sentences, if each subset of $S$ of cardinality $<\kappa$ is satisfiable, then $S$ itself is satisfiable. In fact, this compactness property is needed only for the propositional fragment of $L_{\kappa,\kappa}$ (which I suppose should be called $L_{\kappa,1}$).
Work with a vocabulary that has propositional symbols (= 0-ary predicate symbols) $\hat X$, one for each subset $X$ of $\lambda$. Let $S$ consist of the following sentences of $L_{\kappa,1}$:
$$\widehat{\bigcap_{i\in I}X_i}\iff\bigwedge_{i\in I}\widehat{X_i}$$ for each family $\{X_i:i\in I\}$ of subsets of $\lambda$ where the index set $I$ has cardinality $<\kappa$,
$$ \widehat{\lambda-X}\iff\neg\hat X$$ for each $X\subseteq\lambda$,
and
$$\neg\hat Y$$
for every subset of $\lambda$ of cardinality $<\lambda$.
I claim that, if this set $S$ is satisfiable, say by a truth assignment $v$, then
$$
\mathcal U=\{X\subseteq\lambda:v(\hat X)=\mathtt{true}\}
$$
is a $\kappa$-complete uniform ultrafilter on $\lambda$. Indeed, $\mathcal U$ contains $\lambda$ by the first batch of sentences in $S$ (take $I=\varnothing$), is closed under supersets (take $I=\{0,1\}$ with $X_0\subseteq X_1$), and is closed under intersections of $<\kappa$ sets (take the $X_i$ to be the sets you want to intersect). So $\mathcal U$ is a $\kappa$-complete filter on $\lambda$. It is an ultrafilter by the second batch of sentences in $S$. Finally, it is uniform by the third batch of sentences. So, if $S$ is satisfiable then $\mathcal U$ is as desired.
By strong compactness of $\kappa$, it suffices to show that each subset $S_0\subseteq S$ of cardinality $<\kappa$ is satisfiable. Fix such an $S_0$ and let $B$ be the union of all the $Y$'s such that $\neg\hat Y$ is in $S_0$. Since each such $Y$ has cardinality $<\lambda$, since there are $<\kappa\leq\lambda$ such $Y$'s, and since $\lambda$ is regular, $B$ cannot be all of $\lambda$. Let $\alpha$ be any member of $\lambda-B$ and define a truth assignment $v$ by setting, for each $X\subseteq\lambda$,
$$
v(\hat X)=\mathtt{true} \iff \alpha\in X.
$$
The first two batches of sentences in $S_0$ (in fact in all of $S$) are trivially satisfied by $v$. The third batch of sentences in $S_0$ are satisfied because $\alpha\notin B$, i.e., $\alpha\notin Y$ for any $Y$ such that $(\neg\hat Y)\in S_0$.