zlacker

I think the relationship between model size and training time isn't linear. So if you want a twice bigger model it'll take more resources to train it than two original models.