NASA-IBM Lunar Foundation Model

The TerraMind lineage taken off Earth: a multimodal foundation model for lunar remote sensing pretrained from scratch on SomBench, a geographically partitioned corpus of nearly two million co-registered tile bundles spanning 11 modalities at two spatial scales (1 m and 100 m per pixel). It adapts TerraMind's masked-token architecture with acquisition geometry as explicit context and joint training across both resolutions so one set of weights covers both, and uses FlexiViT patch embeddings so patch size can change without retraining. Trained on 16 H100s in about 1,100 GPU-hours and released under Apache 2.0 with the SomBench data; IBM reports a 22% reduction in ice-detection error over the baseline. Built with NASA's AI4Science group and released 10 September 2026.