You definitely need $M$ to be combinatorial for these types of statements. I believe Lurie has shown that every accessible localization of a presentable infinity category can be expressed as a left Bousfield localization. See chapter 5 section 5 of HTT. He uses *strongly reflective* to mean it comes from an *accessible* localization. Without accessibility, it breaks down, as the next example shows.

**Example**: There are plenty of model categories that don't admit all left Bousfield localizations. The most famous unsolved problem in this vein is cohomological localization for sSet. Very likely, the existence of this localization is equivalent to Vopenka's principle. See work of Casacuberta for evidence to this effect. Right now, it's not known that there is a SET of maps you can invert to do cohomological localization. In Lurie's language, no one can prove it's an accessible localization. Related is the theorem that, if you assume Vopenka's principle, then any left Bousfield localization of any combinatorial model category exists. It's not known if Vopenka can be proven within ZFC, so it would imply a big result in set theory if you found a left Bousfield localization of a combinatorial model category that provably did not exist.

However, as soon as you leave the shelter of combinatorial model categories, there are all sorts of counterexamples. See Chorny's work on class cofibrantly generated model categories. Similarly, for non presentable infinity categories you do not expect all localizations to exist.

If, for some reason, every left Bousfield localization of $M$ was known to exist, then I would expect that any localization of the $\infty$-category of $M$ comes from a localization of $M$. I'd prove this using the universal principle. Since the cofibrations are the same in any left localization, only the weak equivalences matter, and under the assumption about localizations existing, every class of weak equivalences corresponds to a left localization. But this assumption is ridiculously strong, and certainly not necessary for the result you want, as chapter 5 of HTT shows.