NASA and IBM Release Prithvi-Lunar, an Open Model for Analyzing the Moon

On September 10, 2026, IBM and NASA released Prithvi-Lunar, an open-source geospatial foundation model trained on more than a million high-resolution images from the Lunar Reconnaissance Orbiter, with weights published on Hugging Face under an Apache-2.0 license.

What Prithvi-Lunar is and what it was trained on

The model was trained on roughly 2 million image tiles: over 1 million camera images at 1-meter resolution from the Lunar Reconnaissance Orbiter, plus close to 964,000 multispectral images at 100-meter resolution. It's a continuation of the work IBM and NASA have been doing since 2023 on Earth-focused geospatial foundation models under the Prithvi brand, now extended to lunar data.

Where it's available and how it's adapted

The model's weights are published at huggingface.co/ibm-nasa-geospatial under an Apache-2.0 license — permissive enough for unrestricted commercial use with no copyleft strings attached — with the fine-tuning code maintained in a NASA-IMPACT GitHub repository. Adapting it to specific tasks runs through TerraTorch, IBM's framework for fine-tuning geospatial models.

What it's actually useful for

IBM and NASA report that the model can be used to investigate lunar ice deposits, study volcanic formations known as Irregular Mare Patches, and detect craters — outperforming widely-used methods on those three tasks by up to 23%. While the immediate use case is space exploration, the underlying architecture (a multispectral remote-sensing foundation model, fine-tunable with TerraTorch) is the same technical family already used for terrestrial satellite analysis: agriculture, mining, forest monitoring, and parametric insurance.