A new preprint on arXiv details machine-learning models that classify earthquake magnitude using just the first 7 seconds of P-waves from a single vertical-component station. The study curated 7,318 South Asian events split into five Richter classes and benchmarked six approaches, finding a Transformer architecture delivered 76.23 percent standard accuracy and 81.56 percent adaptive accuracy at 4.8 ms latency. This preprint has not yet undergone peer review. The adaptive metric explicitly accounts for boundary uncertainty, a realistic improvement over rigid thresholds used in prior work. Beyond the paper, the approach aligns with Japan's long-operational EEW system (Hoshiba & Aoki, 2015, Earth Planets Space) yet shifts emphasis to attention mechanisms that better handle the long-tail distribution of rare M7+ events. A 2023 USGS study on global early-warning latency (Given et al., Seismological Research Letters) showed each second saved reduces casualties by up to 10 percent in dense urban corridors; scaling the reported 7-second window could therefore deliver actionable alerts to tens of millions across the Himalayan arc within months if integrated with existing networks. Limitations remain: single-station input ignores spatial constraints, high-magnitude samples are scarce, and performance on truly unseen M7+ events is untested outside the catalog. The preprint's emphasis on inference speed nevertheless positions the method as a practical near-term upgrade for regional systems.