IBM and NASA released the NASA-IBM Lunar Foundation Model on September 10, 2026, as an open checkpoint for mapping the Moon. NASA’s science note said the weights are on Hugging Face and the fine-tuning code is on GitHub. A matching IBM wire said the model beat common baselines by up to 23 percent on key surface features.

The model is built to mix instruments that do not share a scale. NASA’s GRAIL gravity maps run at about 20 kilometers per pixel. The Lunar Reconnaissance Orbiter can image boulders at about 1 meter per pixel. IBM and NASA adapted TerraMind, an Earth-observation stack IBM built with the European Space Agency, so those layers sit in one representation. First jobs are uncatalogued craters, volcanic history, and polar ice that could supply water, oxygen, and fuel.

What the checkpoint actually is

The model card describes a ViT-B encoder-decoder trained from scratch on SomBench: about 2 million co-registered lunar tiles across 11 modalities at two scales, LROC NAC near 1 meter per pixel and LROC WAC near 100 meters per pixel. Illumination angles and tile footprints are fed as tokens, because lunar terrain changes more with lighting than with the rock itself. NAC and WAC tiles train in one mixed batch.

IBM’s wire gives three scores against SwinV2-B trained on ImageNet. Ice prospectivity error (RMSE) fell by up to 22 percent. Irregular Mare Patches, the volcanic features, improved about 3 percent with imperfect labels. At context scale of about 100 meters, crater work beat SwinV2-B by nearly 19 percent using half the training data. Downstream heads for craters, IMP segmentation, and ice are posted as separate Hugging Face repos.

“NASA has spent decades building an extraordinary scientific record of the Moon, but collecting data is only part of the job.” Kevin Murphy, chief science data officer and acting chief data and AI officer, NASA

Juan Bernabé-Moreno, director of IBM Research Europe for Ireland and the UK, said the hope is that crews can learn the terrain before they fly. The IBM blog places the model next to Prithvi for Earth and Surya for the Sun. Pretraining code is not in the GitHub release. Fine-tuning and inference are.

Decoded Take

Artemis does not need another pretty Moon mosaic. It needs a way to turn petabytes of mismatched sensors into a map a landing site review can trust. A public foundation model is the right shape for that, because every lab can fine-tune the same backbone instead of training a crater detector from scratch. The numbers that matter are the ice RMSE and the 100-meter crater score, not the 23 percent marketing line. Watch whether NASA flight programs cite this checkpoint in a site-selection package, and whether the GitHub repo grows past fine-tuning into the missing pretraining code. If it stays a research card, the open release is a paper, not a tool for a south-pole sortie.