NASA and IBM open-source a lunar foundation model, Apache-2.0
Trained on ~2M orbiter tiles, it tops the standard baseline by up to 23% on ice and craters, ships a 40TB open dataset, and fine-tunes from a few labels.
Copy markdown
NASA and IBM just open-sourced a Moon model
The NASA–IBM Lunar Foundation Model is out under Apache-2.0 — weights on Hugging Face, full code on GitHub — trained from scratch on roughly 2 million Lunar Reconnaissance Orbiter tiles at 1m and 100m resolution. Apache-2.0 means you can build on it commercially, not just read the paper.
Up to 23% sharper than the go-to baseline
It matches or beats SwinV2-B on every task tested: a 22% error cut on polar ice prospectivity, ~19% better crater detection at 100m, and stronger volcanic-patch mapping — up to 23% overall. That's measurable lift on the exact problems that pick landing sites and hunt water ice.
SomBench: ~40TB of aligned Moon data, free
The training set ships too — SomBench pools ~2 million co-registered tiles across 30+ spatially aligned layers from 9 instruments and 4 missions (LRO, GRAIL, Kaguya), 11 modalities in all, under CC BY 4.0. Even if you skip the model, that's an ML-ready planetary dataset that didn't exist yesterday.
Build this weekend: fine-tune it in two lines
It loads straight through IBM's TerraTorch (BACKBONE_REGISTRY.build(...)), and the model card hits useful accuracy via LoRA, frozen-encoder, or full fine-tunes from small labeled sets. A crater classifier or ice-map tool is a weekend of labeling, not a from-scratch training run on 16 H100s.
The reusable part isn't lunar
The same toolkit and recipe power IBM/NASA's Earth-observation models — so if you touch satellite or geospatial imagery, this doubles as a template: pretrain a ViT-B on aligned multimodal tiles, then adapt with a handful of labels. The Moon is the demo; the pipeline is the takeaway.