Ultra-high field MRI enables sub-millimetre resolution imaging of human brain, allowing to disentangle complex functional circuits across different cortical depths. Segmentation, meant as the partition of MR brain images in multiple anatomical classes, is an essential step in many functional and structural neuroimaging studies. In this work, we design and test CEREBRUM-7T, an optimised end-to-end CNN architecture, that allows to segment a whole 7T T1w MRI brain volume at once, without the need of partitioning it into 2D or 3D tiles. Despite deep learning (DL) methods are recently starting to emerge in 3T literature, to the best of our knowledge, CEREBRUM-7T is the first example of DL architecture directly applied on 7T data. Training is performed in a weakly supervised fashion, since it exploits a ground-truth (GT) with errors. The generated model is able to produce accurate multi-structure segmentation masks on six different classes, in only few seconds. In the experimental part, we show that the proposed solution outperforms the GT it was trained on in segmentation accuracy. For more details, please visit: https://rocknroll87q.github.io/cerebrum7t