IBM and NASA Open Source a Lunar AI Model
IBM and NASA have released the NASA-IBM Lunar Foundation Model as open source, which they call one of the first publicly available foundation models for scientific exploration of the Moon. It is available now and joins IBM's Prithvi family of open models on Hugging Face, which already spans geospatial, weather and heliophysics.
Transcript
IBM and NASA just open sourced an AI model trained on decades of Moon data.
The NASA IBM Lunar Foundation Model is one of the first public models for Moon science. It is on Hugging Face now.
The model predicts where ice may be present in shadowed regions. Ice indicates water and oxygen for a Moon base and rocket fuel.
The paper reports up to 22% less error on likely ice. They also released an open dataset of over 30 layers from nine instruments across four missions.
One caveat. IBM and NASA announced this themselves. Their own paper measured the gains against a single baseline model. This is research and not a mission.
Inboxsmith builds AI phone and email agents for small business. Please like and subscribe for more news.
Sources
Every claim in this video comes from the top ranking coverage of this topic. The claims and where each one came from:
- Every claim in this script comes from IBM's September 10, 2026 newsroom release datelined Yorktown Heights, New York, which announces the open source release of the NASA-IBM Lunar Foundation Model and describes it as one of the first publicly available foundation models for scientific exploration of the Moon, available now and joining IBM's Prithvi family of open foundation models on Hugging Face alongside geospatial, weather and heliophysics models; the release states that sensors have observed the Moon for decades and generated petabytes of data that scientists must otherwise sift through by hand or process with low resolution, task specific machine learning models that are computationally intensive and can lack scientific accuracy; it states that permanently shadowed regions may contain lunar ice below the surface, that the model combines multimodal and multi-resolution observations to predict where ice may be present, and that lunar ice indicates water and oxygen, resources the release calls essential for a future Moon base and for producing rocket fuel for future missions to Mars; the accuracy figures cited in scene three and scene four are the release's own, namely a reduction in error (RMSE) of up to 22% in identifying areas with high potential for lunar ice versus the SwinV2-B (ImageNet) model, nearly 19% better crater performance versus that same baseline at context scale resolution of about 100 meters using half the training data with comparable accuracy and greater efficiency at meter scale, and 3% better capture of the extent of volcanic Irregular Mare Patches using imperfect labels; the release also announces the first open source lunar dataset of its kind, a unified machine learning ready dataset aggregating over 30 spatially aligned layers from nine instruments across four missions, combining tens of thousands of images and maps from NASA's Lunar Reconnaissance Orbiter and NASA's GRAIL mission with complementary data from the Japanese Aerospace Exploration Agency's SELENE/Kaguya, and notes that no publicly available unified dataset of this kind existed before; quotes from Kevin Murphy, chief science data officer and acting chief data and AI officer at NASA Headquarters in Washington, and Juan Bernabe-Moreno, Director of IBM Research Europe, UK and Ireland, appear in the release and informed the framing but were not quoted verbatim in narration. HONESTY CAVEAT: this is IBM and NASA's own announcement about their own model, and the accuracy numbers come from a technical paper titled "Multimodal-Multiresolution Foundation Model for Lunar Remote Sensing" that IBM and NASA authored themselves, benchmarked against a single baseline model, SwinV2-B (ImageNet), with no independent replication cited; the release's top line summary claim of exceeding widely used methods by up to 23% in identifying key geographic features is broader than the specific published gains it details, which are up to 22% on lunar ice error, nearly 19% on context scale craters and 3% on volcanic features; this is a research model release rather than a mission, a landing or a discovery, the model predicts where ice may be present and no ice has been confirmed by it, and YouTube research returned zero usable transcripts for this run because the local yt-dlp build is out of date, so this first party press release is the only source behind the script and no third party reaction, analyst commentary or independent verification was available or used.(IBM and NASA's official press release, "IBM and NASA Release Open-Source AI Model to Support Lunar Exploration", published on newsroom.ibm.com on September 10, 2026)