NASA and IBM open-source lunar foundation model pretrained on 2M multimodal tiles
victormustar · x · 2026-09-19
NASA and IBM have open-sourced the NASA-IBM Lunar Foundation Model on Hugging Face under Apache-2.0, a multimodal foundation model for lunar remote sensing.
Key details:
- A ViT-B encoder–decoder trained from scratch on SomBench: 2 million co-registered lunar tile bundles spanning 11 modalities at two spatial scales (LROC NAC 1 m/px, WAC 100 m/px)
- Built on the TerraMind masked-token recipe with two extensions: acquisition geometry (illumination angles, solar-frame anchors, footprints) tokenized as explicit encoder inputs, and joint mixed-resolution pretraining so one weight set covers a 100× resolution gap
- FlexiViT patch embeddings and modality-wise tokenization allow patch-size changes and adding/dropping modalities at fine-tuning time
USRA contributed planetary science expertise and dataset development. The model supports downstream tasks like lunar ice prospectivity mapping; the paper (arXiv:2504.11171) and fine-tuning code are public.
Related event: NASA and IBM Open-Source Lunar Foundation Model(3 posts)→
More from Research
- Nearly 20 openjev models cataloged as hobbyist preps first community leaderboard — airesearch12 · 2026-09-19
- Mnemos engine builds synaptic-like agent memory via resonant engrams — RileyRalmuto · 2026-09-19
- Shared paper claims AI models have a concept of pain and avoid harm — skolnaja · 2026-09-19
- Steering Vectors as a Softer Way to Limit Reasoning Budgets? — maddie-lovelace · 2026-09-19
- MMLU-Redux paper finds ~6.5% of MMLU questions contain labeling errors — PMinervini · 2026-09-19
- Mathematicians Hate AI. They Can't Quit It — Wired AI · 2026-09-19