NASA-IBM Lunar Foundation Model

The NASA-IBM Lunar Foundation Model is an open-source model for remote sensing of lunar surfaces, released by IBM and NASA on 10 September 2026. It belongs to IBM's Prithvi family of open scientific models, and the Universities Space Research Association, which contributed planetary expertise and evaluation, published its own note on 18 September.

What NASA-IBM Lunar Foundation Model does

• Encoder and decoder: a ViT-B encoder paired with a 12-layer decoder, built from scratch using a masked-token approach adapted from TerraMind.
• Training data: about two million multimodal lunar image bundles across 11 modalities, captured at roughly 1 metre and 100 metres per pixel.
• Fine-tuning: adapts through IBM's open-source TerraTorch toolkit, with LoRA recommended.
• Open release: weights, code, training dataset and benchmarks under Apache 2.0.

Who it works with

Researchers mapping craters, volcanic features and possible lunar ice, using imagery, terrain and map-level context.

Key facts

• Benchmarks: the developers report up to 22% lower error in mapping ice prospectivity than an ImageNet-pretrained SwinV2-B baseline, and nearly 19% better crater detection at about 100-metre resolution using half the training data.
• Limits: the model has no geodetic reference frame, its outputs are uncalibrated, and it is not validated for operational decisions such as landing-site certification.