IBM and NASA Open-Source Prithvi-Lunar: Use It Now

IBM and NASA released Prithvi-Lunar, an open-weight foundation model trained on two million lunar surface tiles, published on HuggingFace under Apache-2.0 alongside SomBench, the first unified ML-ready lunar dataset under CC BY 4.0 on AWS. Prithvi-Lunar beats the SwinV2-B ImageNet baseline on all three core tasks: 22% lower RMSE for ice deposit detection, 19% improvement in 100m-resolution crater detection using half the training data, and 3% better volcanic feature mapping accuracy. IBM states the model lacks geodetic reference validation and cannot certify operational uses such as Artemis landing site safety analysis.

IBM and NASA shipped Prithvi-Lunar today: an open-weight foundation model trained on two million lunar surface tiles, available on HuggingFace under Apache-2.0 right now. Developers can fine-tune it for crater mapping, ice prospecting, and volcanic feature analysis without a NASA research contract. It ships alongside SomBench, the first unified ML-ready lunar dataset — CC BY 4.0, hosted on AWS. The model joins the Prithvi family of open scientific foundation models that IBM and NASA have been building since 2023. What started with Earth observation is now extending to the Moon, and eventually to every scientific domain where closed commercial stacks are the only option. What You Get Access To Right Now Model weights are published at huggingface.co/ibm-nasa-geospatial https://huggingface.co/ibm-nasa-geospatial under Apache-2.0. Fine-tuning runs through TerraTorch https://github.com/IBM/terratorch , IBM’s PyTorch-based geospatial AI library: pip install terratorch The entire fine-tuning pipeline is defined in a single YAML config file — backbone, neck, decoder, and head. The Model Factory assembles it. PyTorch Lightning runs the training loop. Colab notebooks and tutorial examples are in the TerraTorch GitHub repo. The NASA-IMPACT organization https://github.com/NASA-IMPACT/Prithvi-EO-2.0 maintains the downstream adaptation code separately. The Dataset Is Half the Story SomBench is the dataset that makes Prithvi-Lunar possible, and it may be the more lasting contribution. It is the first open, co-registered lunar dataset designed specifically for machine learning — over 30 spatially aligned data layers from nine instruments across four missions: - NASA’s Lunar Reconnaissance Orbiter LRO — cameras, altimeter, thermal, radar, UV - NASA’s GRAIL mission — gravity data - NASA’s Lunar Prospector — hydrogen and neutron flux - JAXA’s SELENE/Kaguya mission That adds up to roughly two million co-registered tile bundles across two resolution tracks: 963,609 tiles at 100 meters per pixel 38 TB and 1,000,113 tiles at 1 meter per pixel 1.4 TB . Every pixel has aligned data across all modalities — no manual alignment work needed. The Benchmark Numbers The baseline for comparison is SwinV2-B pretrained on ImageNet — the standard transfer learning approach most teams would reach for. Prithvi-Lunar beats it on all three core tasks: - Ice deposit detection : 22% lower error RMSE in predicting high-potential ice areas - Crater detection at 100m resolution : 19% improvement, using half the training data - Volcanic feature mapping : 3% better accuracy at lower fine-tuning cost The gains are not surprising once you understand the mechanism. ImageNet pretraining gives a model intuitions built on RGB photographs of Earth. Lunar remote sensing data looks fundamentally different — different spectrum, different surface physics, different imaging geometry. Domain-specific pretraining on lunar data closes that distribution gap. Most developers building geospatial models already understand this intuitively for Earth data; the same logic applies here, just one planet further out. What It Cannot Do IBM is explicit about one hard limit: Prithvi-Lunar is not a scientific-grade generative product. It lacks geodetic reference validation, which means it cannot certify anything for operational use — landing site safety analysis for Artemis missions, for example, cannot be submitted to NASA based on this model’s output alone. It is a research and exploration tool, not a flight-certification instrument. That is the right scope for a v1 open release. But developers integrating this into any workflow with downstream safety implications need to understand that boundary clearly before building. The Pattern Behind the Release Prithvi-Lunar is not an isolated announcement. In roughly 18 months, the IBM and NASA collaboration has shipped Prithvi-EO-2.0 for Earth observation 300M and 600M parameter variants , Prithvi-WxC at 2.3B parameters for weather and climate, and now a lunar model. ESA and IBM co-developed Terramind for multimodal Earth observation on the same TerraTorch infrastructure. One version of Prithvi Geospatial ran in orbit on two satellite platforms — the first AI foundation model deployed in space. The direction is clear: open foundation models as infrastructure for scientific domains where university labs and space startups cannot afford closed commercial stacks. The Moon now has the equivalent of an ImageNet-pretrained backbone. According to the NASA Science team https://science.nasa.gov/science-research/artificial-intelligence-lunar-foundation-model/ , the model is already being used to investigate stable ice locations in permanently shadowed regions — directly relevant to Artemis and commercial lunar lander planning. What to Do Now If you work on geospatial ML, planetary science, or remote sensing, the immediate steps are straightforward. Pull the weights from huggingface.co/ibm-nasa-geospatial https://huggingface.co/ibm-nasa-geospatial , install TerraTorch, and browse the example notebooks. The NASA Science announcement links the technical report. SomBench is on AWS under CC BY 4.0 — downloadable and benchmark-ready without licensing friction. The model is live. The dataset is open. The fine-tuning framework is a pip install away. If your team has ever wanted to build on lunar data and hit a wall because proprietary tools were the only option, that wall is gone today.