Researchers have introduced LibriBrain100, a massive magnetencephalography (MEG) dataset engineered to advance non-invasive neural speech decoding. The dataset provides over 100 hours of high-quality MEG recordings captured while subjects listened to continuous, naturalistic speech. A primary feature of this release is its "depth-first" design, which includes approximately 80 hours of neural data from a single subject—a volume roughly 80 times larger than existing comparable datasets. To complement this depth, the researchers also gathered approximately 40 minutes of data from 32 additional subjects to provide breadth. Evaluating the dataset via a word-classification benchmark, the authors utilized an existing decoding model to achieve state-of-the-art performance, validating the high quality of the recordings. Furthermore, the study demonstrates that supervised fine-tuning using the multi-subject data can significantly mitigate the limitations of small per-subject datasets. To support reproducible research and standardized benchmarking, the authors have released the dataset alongside an open-source Python library for preprocessing and data loading, as well as an open machine-learning competition featuring a public leaderboard. This infrastructure aims to accelerate the development of practical brain-computer interfaces for individuals with severe paralysis.