NASA, IBM Launch Lunar Model as Free Download

NASA and IBM released an open lunar model

NASA Science announced on September 10, 2026, that the agency and IBM Research, working with several academic institutions, had launched the NASA-IBM Lunar Foundation Model, among the first open-source AI models built specifically for lunar science.

The model was trained on an extensive lunar observation dataset curated by IBM and NASA researchers, HPCwire reported on September 10, 2026, and it is aimed at turning decades of complex, multi-instrument records into insights that support a sustained human presence on the Moon.

Kevin Murphy, chief science data officer and acting chief data and AI officer at NASA Headquarters in Washington, said, "NASA has spent decades building an extraordinary scientific record of the Moon, but collecting data is only part of the job." Murphy added that the new model shows what is possible when AI is applied to NASA's petabytes of scientific data.

The weights are posted on Hugging Face, TechRadar reported on September 10, 2026. NASA Science also pointed researchers to a GitHub repository for testing and experimentation.

Silicon Republic reported on September 11, 2026, that the release follows Prithvi, the two organizations' geospatial model, which became the first geospatial foundation model to be deployed in orbit. Simply Wall St reported on September 12, 2026, that the partnership plugs into IBM's broader push to supply AI tools and computing capacity for complex scientific workloads.

Benchmarks show error cuts on ice and craters

A NASA-IBM authored technical paper found the model reduced error (RMSE) in identifying areas with high potential for lunar ice by up to 22% compared with the SwinV2-B (ImageNet) model, HPCwire reported.

On volcanic terrain, the model captured the extent of Irregular Mare Patches 3% better than that same baseline while working from imperfect labels, according to HPCwire, which described the result as comparable accuracy with greater efficiency and lower fine-tuning costs.

The system also maps craters at coarse resolution with 19% better accuracy while using only half as much labeled training data, Tech Times reported on September 11, 2026.

SpaceDaily reported on September 12, 2026, that the output is a model-derived stability map rather than measured ice deposits. Permanently shadowed craters near the poles act as cold traps where water molecules can stay frozen for extremely long periods, SpaceDaily said, and that trapped ice could eventually supply drinking water, oxygen and hydrogen-based rocket propellant.

Tech Times tied the timing to NASA's Artemis program, which is working toward landing astronauts near the lunar south pole.

One dataset stitches nine instruments together

IBM and NASA built SomBench, which Tech Times described as the first open-source machine-learning-ready unified lunar dataset of its kind.

SomBench aggregates more than 30 spatially aligned data layers from nine instruments flown on four missions, principally the Lunar Reconnaissance Orbiter and GRAIL, plus complementary observations from the Japan Aerospace Exploration Agency's SELENE/Kaguya spacecraft, Tech Times reported. Aligning those records by hand has historically required a researcher to convert units and hope no artifacts are introduced, according to Tech Times.

The system was pretrained on roughly two million lunar image tiles, SpaceDaily reported. Silicon Republic reported that the open dataset includes tens of thousands of images and maps drawn from those missions.

Data collected by the Lunar Reconnaissance Orbiter over the past 17 years suited the training run because it covers most of the lunar surface in detail, NASA Science said. Foundation models are pre-trained on vast, unlabeled datasets and generalize across scientific domains through quick fine-tuning, unlike specialized algorithms built from scratch for a single task, according to NASA Science.

KeepTrack counts 2026 orbital launches

KeepTrack published its daily space brief on September 13, 2026, counting 216 orbital launches so far in 2026, 209 of them successful.

Vega-C is set to loft the Sentinel-3C satellite this week, KeepTrack said. Active satellites tracked stood at 15,987 against 14,950 cataloged debris objects, and the next launch on the schedule was a Falcon 9 Block 5 carrying O3b mPower 11-13.

Space-Track cataloged 10 new objects in its latest daily accounting on September 12, 2026, alongside 28,065 updated element sets, KeepTrack reported.

Tron's take

My read is that the transferable part of this story has nothing to do with the Moon. The expensive work IBM and NASA did was curation: aligning more than 30 data layers from nine instruments so one model could learn across them. In my experience, plenty of small and mid-sized businesses that stall on AI stall at that same step, with quote history in one system, tickets in another, and no aligned record connecting them.

My advice is to treat the release as a worked example, not a purchase decision. Nothing here ships a product an owner can deploy this quarter, and I would not move a roadmap because a lunar model landed on Hugging Face. Last quarter's proven capabilities, applied to clean internal data, still pay better than this week's frontier release. That is my reading of the news, not a reported result.

For teams that are evaluating open-weight models, licensing terms and data provenance deserve more scrutiny than benchmark deltas, a point I made when AMD released Instella. XL.net sells managed IT and security assessments; I see nothing in this announcement that calls for either, and I am not pitching one here.