On September 10, 2026, IBM and NASA jointly open-sourced the NASA-IBM Lunar Foundation Model, one of the first foundation models built specifically for lunar science and made publicly available. Model weights are released under Apache-2.0 on Hugging Face; fine-tuning code and training pipeline live in the NASA-IMPACT GitHub repository, with downstream adaptation routed through the TerraTorch framework. The release is paired with SomBench, the first open-source multi-modal, multi-resolution lunar dataset aggregating LRO's 17-year observation record with GRAIL gravity and JAXA SELENE/Kaguya supplementary data — an ML-ready corpus the lunar science community can build on directly.

Architecture and Pretraining: A ViT-B Trained From Scratch With Geometry as a First-Class Input

The architecture is a ViT-B encoder-decoder trained from scratch — 12 layers, 768 dimensions, 12 attention heads — not inherited from ImageNet weights. The training corpus is roughly 2 million spatially co-registered lunar tile bundles from SomBench, spanning 11 modalities at two spatial scales: LRO Narrow Angle Camera imagery at approximately 1 meter per pixel, and Wide Angle Camera imagery at approximately 100 meters per pixel, organized as NAC_HighRes and WAC_LowRes data tracks. The narrow-angle track draws from 1,095 co-registered NAC frames totaling just over 1 million tiles and approximately 1.4 TB; the wide-angle track pulls from 54,080 WAC records totaling just under 964,000 tiles and approximately 38 TB. Both tracks train together in a single mixed-batch loop with one shared set of weights across a 100× resolution gap — a design choice the release emphasizes.

Pretraining ran on 16 H100 GPUs for 150,000 steps at global batch 1,536 in bf16, totaling approximately 1,100 GPU-hours. Modality encoding uses nine modality-specific VQ-VAE tokenizers with FSQ quantization and a DDPM decoder; the objective is cross-entropy over discrete token vocabularies. Per-tile acquisition geometry — illumination angles, solar-frame anchors, tile footprint — is tokenized as explicit encoder inputs. The technical report states this directly: lunar surface appearance is governed more by illumination geometry than by intrinsic surface variation, so treating geometry as a first-class input is more faithful than treating all tiles as i.i.d. images.

Benchmark Results: The Widest Margin Is on Polar Ice Prospectivity

The benchmark table is the most informative part of the release. The headline numbers: on polar lunar ice prospectivity, the model reaches RMSE 0.0293, against 0.0377 for SwinV2-B (ImageNet pretrained) and 0.0397 for a random-initialization control — ice prospectivity is flagged in the report as the widest-margin task. On Irregular Mare Patch (IMP) segmentation, IoU1 hits 0.5709 versus 0.5687 for ConvNeXtV2-B; the margin is small but the model maintains the lead under noisy labels. On WAC Robbins crater detection with LoRA fine-tuning, mAP reaches 0.2581 against 0.2420 for SwinV2-B. IBM's published framing is "up to 22% lower RMSE on ice prospectivity" and "nearly 19% better at roughly 100-meter context scale using half the training data" — these are the maximum-margin figures from the same technical paper, not averages, and the authors themselves footnote that NAC crater and IMP margins are within seed-spread and should be read as "comparable" rather than significantly better.

Boundaries and Positioning: Accelerating Hypothesis Generation, Not Last-Mile Mission Decisions

The boundaries need to be explicit. The model card self-states three limits: it is not a science-grade generative model; it carries no geodetic reference frame; and it is not validated for operational decisions such as landing-site certification or hazard clearance. Ice-prospectivity outputs regress a knowledge-driven fuzzy overlay rather than measured ice. The SomBench dataset card also notes heterogeneous spatial resolution and tens-of-meters absolute geolocation uncertainty. In short: this is a pretrained base for accelerating hypothesis generation and narrowing the search space — not a last-mile decision tool for mission operations.

Prithvi Series Extension: From Earth to the Moon, the "Base Plus Data Plus Tooling" Troika

The model joins the Prithvi family of open foundation models from IBM-NASA, previously covering geospatial, weather, and heliophysics, now extended to the Moon. NASA Chief Science Data Officer Kevin Murphy and IBM Research Europe UK/Ireland Director Juan Bernabe-Moreno both frame the release around "starting from a shared base instead of building a new algorithm for every scientific question" — a consistent narrative across the Prithvi series. The work was supported by NASA under Award No. 80MSFC25M0084. The full release is live on Hugging Face (nasa-ibm-ai4science/NASA-IBM-Lunar-Foundation-Model) and the NASA-IMPACT GitHub repository.

So What: Why This Release Is Worth Its Own Write-Up

The Prithvi series extends the "domain foundation model + publicly co-registered dataset" pattern from Earth to the Moon. Laboratories working on lunar surface science no longer need to train a ViT from scratch or write glue code to align LROC, GRAIL, and Kaguya data across resolution tiers — they can fine-tune or LoRA-adapt the encoder to their specific task. This combination of "dataset + pretrained base + downstream tooling, all open-source" lands in the context of NASA's return-to-the-Moon program and effectively pushes lunar remote sensing from a closed-door effort by a few teams to something any Earth-based CV group can pick up. That is why IBM frames it as "one of the first publicly available foundation models for scientific exploration of the Moon" — the emphasis is not "largest" or "strongest" but "public, modifiable, usable." The release also partially answers recurring scientific community questions about whether AI-for-Science tools can actually be open: weights on Apache-2.0, code on GitHub, and data on CC BY 4.0 all together is the furthest a comparable Earth-science model release has gone.

References: