First of all, regardless of speculative comments in the original BFSS paper, different superselection sectors – different parts of the Hilbert space specified by the "background" (at least at infinity) – require different matrix model descriptions.
The matrix model is known for $T^p$ with any radii; for $p\leq 3$, the theory is a super Yang-Mills theory on the dual torus, for $p=4$, the 4+1-dimensional Yang-Mills on $T^4$ is non-renormalizable but automatically completed to the $d=6$ (2,0) superconformal field theory, and for $p=5$, one gets little string theory on a $T^5$. There exists no non-gravitational matrix model description for $p\geq 6$ although such models would be interesting as they would exhibit exceptional Lie symmetries.
The radii may be anything; but the 11-dimensional Planck scale is the only natural "unit of length" in the 11-dimensional M-theory, so whatever the radii are, they are naturally expressed as multiples of the Planck length. The numerical multiples are pure numbers; if you invert them, you get the radii of the dual torus in the Yang-Mills matrix model expressed in its natural length scale given by the Yang-Mills scale. (For $p=3$, the theory is conformal, so some parameters get translated to the dimensionless gauge coupling. For $p=4$, the theory in $d=6$ is also conformal but there's one extra dimension whose radius defines a preferred scale. The little string theory for $p=5$ has its own "string scale".)
But the matrix model works even if the radius is much longer or much shorter than the 11-dimensional Planck scale. For example, M-theory on a parameterically short $S^1$ circle produces weakly coupled type IIA string theory whose matrix model, Yang-Mills on $S^1\times R$, is also known and may be shown to describe the IIA strings via the so-called matrix string theory, including all the degrees of freedom on the strings of any length, their interactions, and D-branes.
