Image inpainting of brain tumour MRI volumes has the potential to correct image artefacts, enable novel data augmentation methods and increase compatibility with automated image preprocessing methods. However, achieving 3D anatomical consistency and radiological fidelity remains a significant challenge, especially in the absence of large, annotated datasets. The BraTS 2025 dataset included T1-weighted brain MRI data from 1251 patients for training and 251 for validation. In our solution, we train an unconditional 3D diffusion transformer model to generate synthetic brain MRI data, operating directly on the high-dimensional pixel space ( \(240\times 240\times 155\) ). To ensure computationally efficient training, we utilise several architectural choices (Hourglass Transformer, efficient attention variants). The models are trained with affine data augmentations, a sigmoid noise schedule and Min-SNR loss weighting. Importantly, our solution does not use any inpainting-specific training. During inference, in the validation set, our models achieve a mean squared error (MSE) of 0.013, peak-signal-to-noise ratio (PSNR) of 20.23 and a structural similarity index measure (SSIM) of 0.775. Qualitatively, the inpainted regions are anatomically consistent with the surrounding tissue. In conclusion, our solution shows the feasibility of using a fully 3D diffusion model to address the task of healthy tissue inpainting, without inpainting-specific training.