I would be interested in how CLIP performs post quantization. This should be relatively simple to test via ImageNet zero-shot top-1 acc as one example metric.