I believe that what you want to do is possible in principle, but I would have concerns that you would be updating too many voxels per frame and that you wouldn't be able to regenerate the mesh fast enough.
There is a game in development called 'Minecraft' which also uses a volumetric representation of the world, and they do have this concept of water as a voxel type. They apply basic physics so that the water will flow into open spaces. However, it's very slow and not realistic. You can see it in action here:
http://www.youtube.com/watch?v=6mxeKZzt5MUHere's one on a larger scale:
http://www.youtube.com/watch?v=JXmSUKrFemc (starts at 40 seconds in)
Is this basically the kind of effect you want to achieve?
From their various videos I get the impression that their engine is a lot slower that PolyVox/Thermite3D (though maybe this is just the artstic effect they aim for), so I think that we could do a lot better. However, I still doubt if it is possible to perform simulation on a large volume in real time.
The problem is that Thermite3D works best when local modifications (such as explosions) are performed because only a small part of the surface has to be rebuilt. With flowing water I can imagine that the whole volume (or most of it) is changing every frame. If you want to regenerate the whole surface every frame then I would estimate that 64x64x64 voxels is the largest volume size you could use.
That said, if you only have a finite amount of water in a level then it can't be everywhere at once, and maybe this keeps the surface regeneration to a minimum.
Thermite3D is not in a mature enough state to prototype this, but PolyVox is. You'd essentially have to write a simple viewer app (based on the OpenGL example perhaps) and then write the code to set the voxels to rock/air/water based on your own rules.
However, overall I believe the best approach for what you are trying to do would instead be to represent the water as particles rather than voxels. I would use an existing physics engines which supports fluid dynamics, pass the terrain geometry from PolyVox into the physics engine, and then render the resulting partcles using metaballs or some other approach. I think your water would look better and move more realistically this way.