Huggingface or an academic effort could also improve the efficiency of few/one shot learning, or simply a crowdsourced model build seems within reach.
I guess to run the model you need a large machine too.