AI Potluck
Back to Gap Map Model components / Scientific AI models

NASA-IBM Lunar Foundation Model

NASA IMPACT
open weights / Overall score: 3.1

A ViT-B encoder-decoder pretrained from scratch on SomBench, roughly two million co-registered lunar tile bundles spanning 11 modalities at two spatial scales, LROC NAC at about 1 m/px and LROC WAC at about 100 m/px. It adapts TerraMind's masked-token recipe to the Moon by tokenizing acquisition geometry as explicit context and training both resolution families in one mixed-batch loop, and downstream adaptation runs through TerraTorch with runnable configs for crater detection, Irregular Mare Patch segmentation and ice prospectivity.

GitHub's license endpoint reports NOASSERTION, but the repository LICENSE body is Apache-2.0 and both READMEs and the model card say the same. The companion repository states that pretraining code is not included, and the published SomBench dataset repo is a sample of the corpus rather than the corpus. Verified 2026-09-10 via the Hugging Face API, the model card, and the GitHub repository.

Openness

3 high confidence
3.0
weights
open(ungated Apache-2.0 backbone checkpoint plus nine modality tokenizer checkpoints)
data
documented-not-released(SomBench composition given tile by tile
code
partial(TerraTorch fine-tuning integration and configs
license
Apache-2.0(OSI)

An unusually complete release for a government-lab model, and it still stops short of the top rungs. The corpus is described down to per-modality counts and split rules, but what ships on the Hub is a sample, and the fine-tuning repository says in one line that pretraining code is not included.

Adoption

1 high confidence
1.0

Three downloads in the trailing 30 days against a card whose cited creation timestamp puts it barely over a week old, so this is a first-week number rather than a settled one.

Capability

4 high confidence
4.0

Three of the four attributes are stated outright. The fourth is refused by the publisher rather than missing - the card calls any-to-any generation a qualitative probe and says the model is not a scientific-grade generative product, which is the category's own test for what a generated output has to be.

Verified 2026-09-10