I think they will not all be CM though, at least for $n = 2$.
I can compute an explicit non-CM weight $(1,3)$ Hilbert modular form $f$ over $F = \mathbb{Q}(\sqrt{5})$. If I consider the $\ell$-adic Galois representation $\rho_f \colon G_F \rightarrow \mathrm{GL}_2(\overline{\mathbb{Q}_\ell})$ for some prime $\ell$ split in $F$, say with $v, v' | \ell$, then the Hodge-Tate weights of $G_{F_v}$ and $G_{F_{v'}}$ are (WLOG) $(1,1)$ and $(0, 2)$.
The tensor induction $\mathrm{TInd}_{G_{\mathbb{F}}}^{G_{\mathbb{Q}}} \rho_f$ will then be a $4$-dimensional strongly irreducible geometric $G_{\mathbb{Q}}$-representation with Hodge-Tate weights $(1,1,3,3)$. Twisting the tensor induction by the inverse cyclotomic character, I can get a $4$-dimensional geometric $G_{\mathbb{Q}}$-representation $\xi$ with Hodge-Tate weights $(0,0,2,2)$.
For the specific form $f$ that I started with, the coefficients satisfy some nice symmetries, and the Frobenius traces of $\xi$ appear to all be valued in $\mathbb{Q}$.
Conjecturally, I think this $\xi$ should be the $\ell$-adic realization of some motive $M$ over $\mathbb{Q}$ with coefficients in $\mathbb{Q}$ and Hodge vector $(2,0,2)$. This motive $M$ shouldn't be CM since $f$ wasn't CM - I think $M$ should have Mumford-Tate Lie algebra $\mathfrak{sl_2}^{\oplus 2}$ from the construction.
