I don't know if they have a theory of mind, but some animals absolutely have very complex behavioural prediction models for humans and other animals. I swear my last dog knew whether I intended to go outside based solely on how I stood up from the couch.

This can be explained as associative learning, without needing to reach for ToM, which is still no mean feat.

I believe it's been shown as a general capability in [corvids](https://www.semanticscholar.org/paper/Ravens-attribute-visua...), great apes and maybe possibly in pigs (and goats). It's sort of game theoretic, as in "I know that you know that I stash food here so I'm going to pretend to hide the food here". I've never read anything that goes beyond two levels of "I knows" but it can be sophisticated, such as in the linked paper on Ravens where there is line of sight through a peephole that can be open or closed.

Sophisticated ToM is likely rare in the animal kingdom, not usually going beyond programmed responses to being observed and possibly empathizing with conspecific suffering.

In humans, it's a much more sophisticated process of inferring motivations, anticipating moves and a number of similar inferential capabilities plus abductive reasoning.