Here's one approach that I expect should give you a multiplicative-factor approximation, with polynomial time running time.
Let $L$ be a regular language that is a subset of $\{0,1\}^n$, e.g., $L=L(M) \cap \{0,1\}^n$. We will try to compute the approximate size of $L$.
At a high level, our approach to approximate $|L|$ will look something like this:
Pick a fraction $p$, where $0<p<1$.
Choose a regular language $R$ such that, roughly speaking, $R$ is a random subset of $\{0,1\}^n$ of size approximately $p2^n$ (i.e., $|R|\approx p2^n$).
Check whether $L \cap R$ is non-empty. Note that this check can be done in polynomial time.
Repeatedly perform steps 1-3 for various values of $p$. This gives you some information that will let you approximate $|L|$.
In particular, if $|L|=m$, then we'd expect
$$\Pr[L \cap R = \emptyset] = (1-p)^m \approx e^{-pm}.$$
So, if you happen to choose $p=1/m$ and repeat steps 1-3 a bunch of times, you should expect to see an empty intersection about 37% of the time. If you see an empty intersection significantly more often then that, then increase $p$ and try again. If you see an empty intersection significantly less often then that, you might decrease $p$ and try again.
In this way, using something like binary search, you should be able to approximate $|L|$ to within a multiplicative approximation factor.
You'll still need to pick some way to choose $R$ so that it is regular but also behaves like a random subset. There are many possibilities, but one might good way might be to choose a random 2-universal hash $h:\{0,1\}^m \to \{0,1,2,\dots,k-1\}$, pick $y \in \{0,1,\dots,k-1\}$ randomly, and let $R=\{x \in \{0,1\}^n : h(x)=y\}$. Choosing $k=\lceil 1/p \rceil$ gives you a random set $R$ of approximately the right size, and because $h$ is 2-universal, all of the mathematics above should work out properly.
This should solve your problem for the case where all strings in the NFA have the same length, say $n$. If they have varying lengths, then you can handle each possible length separately. Since $M$ is acyclic, the maximum lengths of any string in $L(M){}$ is at most the number of states in $M$, so this does not increase the runtime too much.
(This construction might remind you of the Valiant-Vazirani theorem about unambiguous SAT.)