NASA-IBM Lunar Foundation Model

The NASA-IBM Lunar Foundation Model is an open-source AI model for lunar remote sensing, released by IBM and NASA on 10 September 2026. IBM places it in the Prithvi family of open scientific foundation models, alongside its geospatial, weather and heliophysics models. The Universities Space Research Association, which contributed planetary science expertise, dataset work and evaluation, published its own release about the model on 18 September.

It is a ViT-B encoder with a 12-layer decoder, trained from scratch with a masked-token method adapted from the TerraMind Earth-observation model. Pretraining used SomBench, about two million co-registered lunar tile bundles covering 11 modalities at two scales: roughly 1 metre per pixel from the Lunar Reconnaissance Orbiter's narrow-angle camera and 100 metres per pixel from its wide-angle camera. Inputs include imagery, terrain, slope and illumination geometry, plus map-level context such as thermal, radar, mineralogy, gravity and hydrogen data. Fine-tuning runs through IBM's open-source TerraTorch toolkit, and LoRA adaptation is recommended. The weights and code are released under Apache 2.0, together with the training dataset and benchmarks.

In the developers' own technical paper, the model reduced error in mapping lunar ice prospectivity by up to 22% against an ImageNet-pretrained SwinV2-B baseline, and improved crater detection at about 100-metre resolution by nearly 19% using half the training data.

The model card is explicit about its limits. It is a research representation rather than a scientific-grade generator: it keeps no geodetic reference frame, its generated values are not calibrated, its ice output reproduces a prospectivity map rather than measured ice, and it has not been validated for operational decisions such as landing-site certification.