NASA and IBM Open Source a Foundation Model for Mapping the Moon

IBM and NASA have shipped another entry in their Prithvi family of open source foundation models, this time pointing the tech at the Moon. The NASA-IBM Lunar Foundation Model is now live on Hugging Face, trained on a curated lunar observation dataset the two organizations spent time assembling from multiple missions and instruments.

The pitch here isn't just "AI for space, cool." It's a real engineering problem: lunar data comes from different instruments, different resolutions, different viewing angles, and none of it plays nicely together by default. NASA and IBM claim this is the first model to integrate observations across modalities (different data formats and sensor types) and spatial scales into a single unified model. That matters because the alternative, up to now, has been scientists manually cross-referencing maps and imagery, or relying on lower-resolution ML models that couldn't handle the full mix of inputs.

Practically, the model is aimed at three specific science tasks:

• Finding lunar ice deposits in permanently shadowed regions, some of the hardest terrain on the Moon to observe directly. Ice matters because it implies water and oxygen, both critical resources for any future crewed missions that don't want to haul every drop of water from Earth.
• Analyzing Irregular Mare Patches, volcanic features that give clues about the Moon's geological history.
• Identifying and classifying craters at scale, rather than one at a time by eye.

For the ice-hunting use case specifically, the model combines multimodal and multi-resolution observations to better predict where ice is likely hiding, which is a meaningfully harder problem than just running object detection on a single image layer.

The Dataset Behind the Model

The model is only half the story. Alongside it, NASA and IBM released an open-source dataset built from more than 30 spatially-aligned layers, pulling from nine instruments across four separate missions. That's tens of thousands of images and maps, all lined up so they describe the same patches of lunar surface across different geophysical properties. Getting that spatial alignment right, so a pixel in one layer actually corresponds to the same physical location in another, is the unglamorous but essential groundwork that makes a multimodal model like this trainable in the first place.

"NASA has spent decades building an extraordinary scientific record of the Moon, but collecting data is only part of the job," said Kevin Murphy, the agency's chief science data officer. "We also have to make data easier for scientists to explore and use."

IBM's framing leans on the same idea. Juan Bernabe-Moreno, the company's director of research for Europe, described the model as giving scientists "a foundation to explore the Moon at scale, connecting observations across instruments, revealing patterns that are difficult to see in isolation, and providing an open platform the global research community can build on."

This isn't IBM and NASA's first rodeo. The Prithvi lineage started in 2023 with a foundation model for labeling satellite imagery of Earth. That was followed a year later by a climate model built for weather prediction, and last year by Surya, a model trained to forecast solar flares that could mess with satellites and spacecraft. The Lunar Foundation Model slots into that same family, applying the same general approach, pretrain on a huge pile of aligned scientific observations, then let researchers fine-tune for specific tasks, to a new domain.

One thing notably absent from the announcement: a parameter count. NASA and IBM haven't disclosed the model's exact size. When asked what hardware someone would actually need to run or fine-tune it, an IBM spokesperson gave the kind of "it depends" answer that will sound familiar to anyone who's tried to spec out infrastructure for a foundation model:

"Hardware needs will depend on the application, the size and number of inputs, and whether they're running predictions or further training the model. As a rule of thumb, most of our fine-tuning experiments were conducted using Nvidia A100 GPUs."

They added that smaller-scale experiments and inference workloads "may be possible on more modest hardware," though the exact requirements vary by task. Translation: if you're doing serious fine-tuning across the full multimodal dataset, budget for A100-class compute. If you just want to poke at inference on a subset of the data, you've got more room to improvise.

Key Takeaways

• The NASA-IBM Lunar Foundation Model is open source and available now on Hugging Face, built to unify multimodal, multi-resolution lunar observations in a single model rather than forcing researchers to stitch data sources together by hand.
• It's paired with an open dataset of 30-plus spatially-aligned layers pulled from nine instruments across four missions, tens of thousands of images and maps deep.
• Target science use cases include hunting for lunar ice in permanently shadowed craters, studying Irregular Mare Patches, and automating crater identification and classification.
• No official parameter count has been released. IBM says most fine-tuning work used Nvidia A100 GPUs, though lighter inference tasks may run on more modest hardware depending on input size and task complexity.
• It's the latest release in IBM and NASA's Prithvi foundation model family, following Earth imagery labeling in 2023, a climate model in 2024, and the solar-flare-predicting Surya model in 2025.

For developers and researchers who've been curious about foundation models applied outside the usual text-and-image domains, this is a solid one to pull down and poke at. It's open weights, it's got a real dataset behind it, and unlike a lot of "AI for science" announcements, there's an actual model card on Hugging Face you can go read right now.