A team is predicting house prices from 12 features. They have 5,000 training houses and a separate set of 1,000 validation houses that are never used for training. They train four networks of increasing size, each for the same long run, and record the final losses. The loss is mean squared error on standardised prices, so the numbers can be compared directly:
| model | weights | training loss | validation loss |
|---|---|---|---|
| A | 400 | 0.68 | 0.70 |
| B | 9,000 | 0.31 | 0.34 |
| C | 120,000 | 0.12 | 0.27 |
| D | 2,500,000 | 0.01 | 0.55 |
Which of these four trained models should the team deploy to price houses it has never seen?
Select all that apply.