We can model many "reality" with the classical methods good enough, using Deep Learning and Tensor Network based methods, see for instance - https://quantumfrontiers.com/author/roundsphereblog/ - https://www.quantamagazine.org/key-chemistry-question-answer...

When it comes to many of these systems relevant real world (e.g. quantum chemistry), classical heuristic-based approaches are already successful enough. For instance, you can run one of the simulations in Garnet Chan paper in a 10-15 minutes using some machine similar to DGX Spark to simulate FeMo-cofactor model within accepted quantum chemistry precisions.

I believe its biggest application will be to explore some areas in quantum information (e.g. quantum coherence), all the practical applications will be minor.

If quantum computers turn out to be able to do that, the api would be very similar to what dwave has today.

That is one will use normal computer to reformulate the task into a predefined form, submit that to quantum computer, and then use normal computer again to process the results.

Quantum computers can't model reality in general. E.g. we're not going to be doing weather prediction or traffic modeling with quantum computers.