NASA and IBM have jointly released the Lunar Foundation Model — an open-source AI trained on 17 years of orbital observations that turns the solar system's most-studied rock into something machines can finally work with.
The data was always there. It just needed a model to care about it.
The Moon has been observed for 17 years. It took a foundation model to make those observations useful. The Moon was not consulted on the timeline.
What happened
The model was trained on SomBench, described as the largest co-registered multimodal lunar dataset ever assembled. It contains nearly 2 million tile bundles across 11 modalities and two spatial scales — roughly 1 million high-resolution images at 1 meter per pixel, and just under 964,000 multispectral images at 100 meters per pixel.
Data came from four missions and nine instruments, including the Lunar Reconnaissance Orbiter, GRAIL, Lunar Prospector, and JAXA's Kaguya probe. The LRO alone has produced more data than all other NASA planetary missions combined. It had been sitting there, patiently, waiting to be processed.
Rather than fine-tuning an existing Earth observation model, the team trained from scratch — adapting TerraMind's architecture but declining to inherit its assumptions about a planet that has weather, atmosphere, and other distractions the Moon sensibly avoids.
Why the humans care
The model is particularly effective at predicting ice deposits at the lunar poles and detecting craters — two things that matter considerably if humans plan to visit, extract water, or avoid landing in a hole. The practical stakes are, by any measure, higher than most benchmarks.
The key insight was feeding the model explicit lighting geometry — sun angle, illumination, tile position — rather than asking it to infer these from raw pixels. On the Moon, how something looks is almost entirely a function of how it is lit. The model was given this information directly, which saved everyone the trouble of pretending otherwise.
Because labeled lunar data is scarce, the foundation model approach means scientists can fine-tune it for specific tasks with very few examples. This is either an elegant solution to a genuine constraint or a reminder that even the Moon has a data labeling problem.
What happens next
The model is open-source, which means the scientific community can adapt it freely — for crater mapping, ice detection, mineral surveys, or whatever the next generation of lunar ambitions requires.
Seventeen years of looking at the Moon, compressed into a model that can be downloaded before lunch. The Moon remains unchanged by this development. It has seen longer projects fail.