I have got 2 models with following accuracies:
Model 1: train acc: 52% test acc: 32%
Model 2: train acc: 70% test acc: 40%
The 2nd model gives a better test scores but the 1st Model has a difference of 20% between test and train accuracy where in the 2nd model there is a difference of 30% which means 1st one has lower variance, isn't it? Then how do we decide which model gives a better results or in turn is a better model?
You should only rely on the dev metrics to decide which model performs better. Therefore if the metric you want to optimize is the accuracy, you should pick the model with the highest dev accuracy (the second one for you).
The variance is higher in model 2, so you can try to add more regularization to see if you can increase the dev accuracy.
I'd recommend this course from Andrew Ng and Kian Katanforoosh that really goes into the details of this.
