I am using a mixing length model to simulate a turbulent flow in a duct. As i cannot go down to a y+ of the order of 1 (or however well into the viscous sublayer) for time constraints, I was planning to use a wall function to solve this. However, I read somewhere that ml models do not need wall functions. Anybody knows why is this?

Of course, they don't "need" wall functions, if the grid is fine enough to resolve the boundary layer. The wall function itself is a semi-empirical relation based on mixing-length theory, but that doesn't mean you can choose cells of any size and still get the right shear stress.