NASA and IBM Open-Source a Moon Model That Warns Against Trusting It
A.I. / news
NASA and IBM Open-Source a Moon Model That Warns Against Trusting It
The model cuts ice-detection error by 22 percent against a standard baseline, but its own documentation says its ice maps are not measured ice and it is not fit for landing-site decisions.

NASA and IBM released an open-source foundation model for mapping the Moon on Sept. 10, and the model's own documentation warns against using it for the decisions it was built to inform.
The NASA-IBM Lunar Foundation Model is trained on more than 30 spatially aligned data layers drawn from nine instruments across four lunar missions, including NASA's Lunar Reconnaissance Orbiter and GRAIL, and Japan's SELENE/Kaguya, according to IBM Research. The model is meant to help scientists search for polar ice, catalog small craters and trace the Moon's volcanic history from data that previously required separate, task-specific tools for each job.
"We also have to make data easier for scientists to explore and use," said Kevin Murphy, NASA's chief science data officer and acting chief data and AI officer, in the IBM announcement. Michael Barker, a NASA lunar topography expert, co-led the project with Juan Bernabé-Moreno, director of IBM Research Europe for the UK and Ireland.
Where it beats a plain baseline, and by how much
Against a SwinV2 transformer baseline, the model reduced error in ice-prospecting predictions by 22 percent, IBM said. On crater detection at 100 meters per pixel, it beat the baseline by nearly 19 percent while training on half as much data. It matched the baseline's performance on crater detection at 1 meter per pixel, and beat it by 3 percent on mapping volcanic features.
| Task | Result vs. SwinV2-B baseline |
|---|---|
| Ice-prospecting error (RMSE) | 22% lower |
| Crater detection, 100 m/px | ~19% higher, on half the data |
| Crater detection, 1 m/px | Matched |
| Volcanic feature mapping | 3% higher |
Ninety percent of it is frozen
The model is adapted from TerraMind, an IBM-ESA Earth observation model, using lightweight low-rank adapters while keeping 90 percent of the original weights frozen, according to its Hugging Face model card. It is a ViT-B encoder-decoder with 768 dimensions, 12 layers and 12 attention heads, taking 256-by-256-pixel image patches across 11 modalities, and it was trained on 16 Nvidia H100 GPUs for 150,000 steps, or roughly 1,100 GPU-hours. It joins IBM's Prithvi line of open foundation models, which already covers Earth observation and heliophysics, and is licensed under Apache 2.0. The weights have drawn 2,651 downloads on Hugging Face in the past month.
What its own card says it cannot do
The model card describes the release as "not a scientific-grade generative product." It says the model's outputs "are not calibrated predictions" and carry no geodetic reference frame, and states plainly that the model is "not validated for operational decisions" such as certifying a landing site. The ice-prospecting output, the card says, is "a knowledge-driven fuzzy overlay" rather than a direct measurement of ice. Evaluation was limited to lunar products in a dataset called SomBench, which the card does not describe as independently audited.
Writing that kind of limitation directly into a model card is not universal practice this month: this site's earlier report on PrismML's Ternary Bonsai 2 found a 75 percent score that hid a shortfall its own documentation never mentioned. NASA and IBM have not said whether a version validated for mission planning, following Alibaba's Qwen3.8-27B release and other open-weight drops this month, is in development, or when the next model in the Prithvi family will arrive.
Sources
More in A.I.
- 01Anthropic's Levent Alpöge Claims a Claude-Assisted Riemann Proof, and Lean Checks Are InImperial's Kevin Buzzard and MIT's Andrew Sutherland have examined the formalization, according to Scientific American. No journal review has been reported.
- 02Thinking Machines Ships Inkling 2 Mini, a 26.4B Apache 2.0 Model Sized for One DGX SparkThe model card lists 3.4B active parameters, a 1M-token window and a 20GB checkpoint, but does not say who ran most of its benchmark rows.
- 03Anthropic Closed Its Debut at $97.17, 17% Over the $83 IPO Price, as Nasdaq Fell 2%The $81.34 billion offering opened 34% higher, then gave back most of the gain while Nvidia, AMD and Micron dropped.
- 04Google's Gemini Tool Retrieval Preview Adds 26 Tokens Per Idle Tool and Publishes No Savings FigureAnthropic and OpenAI each reported token reductions for the same defer_loading flag. Google's November 19 post reports none.