<p>Until recently, our pure Python, primitive equation ocean model Veros&#160;<br>has been about 1.5x slower than a corresponding Fortran implementation.&#160;<br>But thanks to a thriving scientific and machine learning library&#160;<br>ecosystem, tremendous speed-ups on GPU, and to a lesser degree CPU, are&#160;<br>within reach. Leveraging Google's JAX library, we find that our Python&#160;<br>model code can reach a 2-5 times higher energy efficiency on GPU&#160;<br>compared to a traditional Fortran model.</p><p>Therefore, we propose a new generation of geophysical models: One that&#160;<br>combines high-level abstractions and user friendliness on one hand, and&#160;<br>that leverages modern developments in high-performance computing and&#160;<br>machine learning research on the other hand.</p><p>We discuss what there is to gain from building models in high-level&#160;<br>programming languages, what we have achieved in Veros, and where we see&#160;<br>the modelling community heading in the future.</p>