Vision models are way smaller and run decently on these things.