Tensil founder here: we take fully trained models, training doesn't happen on our chips. The model architecture is hardened but the parameters can be reprogrammed.