Also Read: Development: Exploring Application Security:
IBM and NASA Just Open-Sourced an AI Model of the Moon
Here's a nice break from antitrust probes and zero-days. On Thursday, IBM and NASA released the NASA-IBM Lunar Foundation Model, an open-source AI model built to help scientists make sense of decades of lunar data.
Also Read: News: Apple Finally Folds:
The problem it solves
We've been photographing, scanning and mapping the Moon for a very long time. The result is petabytes of data from lots of different instruments, captured at different resolutions and measuring different things. Traditionally, scientists either sift through it by hand or build narrow machine learning models that only handle one task.
A foundation model flips that around. You train one large model on everything, then fine-tune it for whatever question you're trying to answer.
Also Read: Commentify Pro: Headless Comments for Laravel
What it was trained on
The training dataset, called SomBench, brings together tens of thousands of images and maps from NASA's Lunar Reconnaissance Orbiter and GRAIL missions, plus data from Japan's SELENE/Kaguya mission. Put together, that gives the model a view of both the lunar surface and what lies beneath it.
What it can actually do
Also Read: Laravel: A Self-Hosted Disqus
IBM and NASA are highlighting three main uses.
Finding ice. Permanently shadowed craters near the lunar poles may hide water ice, which future missions could turn into drinking water, oxygen and even rocket fuel. IBM says the model cut the error in spotting high-potential ice areas by more than a fifth compared with SwinV2-B, a widely used general-purpose vision model.
Mapping craters. This is how NASA picks safe landing sites and avoids steep slopes and boulders. At metre-scale resolution, the new model performs about as well as existing top models, but it's cheaper and more efficient to fine-tune.
Volcanic history. Here the gains are smaller, around 3%, partly because the training labels themselves aren't perfect.
It's worth being clear that the ice result is the headline number. The other improvements are more modest, but lower fine-tuning costs still count for a lot, especially for smaller research teams.
Why developers should care
Also Read: Chrome 153 Fixes 7th Zero-Day of 2026: Update Now
The model weights are published on Hugging Face, and the fine-tuning code sits in a NASA-IMPACT repository on GitHub, with adaptation handled through the TerraTorch toolkit. In practice, that means university labs, students and small space startups can build on it without paying for a closed commercial system.
It also joins IBM and NASA's Prithvi family of open models, which already covers Earth observation, weather and heliophysics. The long-term goal is simple: stop building a brand-new model from scratch for every scientific question.
Our take
With space agencies pushing towards a sustained human presence on the Moon, knowing where the ice is and where it's safe to land matters a great deal. This isn't the flashiest release of the week, but open tools like this are exactly how smaller teams get a seat at the table.
There's also a small irony in the timing. The model is launching on Hugging Face just as Nvidia has agreed to buy the platform. Open science, meet big business.
