IBM and NASA announced the open-source release of the NASA-IBM Lunar Foundation Model on September 10, 2026, describing it as one of the first publicly available foundation models built specifically for scientific exploration of the Moon’s surface. Alongside the model, the two organizations released SomBench, a co-registered lunar dataset that aggregates more than 30 spatially aligned data layers drawn from nine instruments across four separate lunar missions - the first open-source dataset of its kind unifying that much multi-instrument lunar observation data in one place.
The model is multimodal and multi-resolution, meaning it was built to reason across different types of lunar remote-sensing data captured at different scales rather than a single imaging source. In benchmark comparisons described in the release, it cut error rates by up to 22 percent (measured by RMSE) in identifying potential ice deposits, improved capture of volcanic surface features by roughly 3 percent, and outperformed a SwinV2-B baseline by about 19 percent at detecting craters at context scale. NASA’s Kevin Murphy framed the release around a data problem rather than a collection problem: “NASA has spent decades building an extraordinary scientific record of the Moon, but collecting data is only part of the job.” IBM Research Europe’s Juan Bernabe-Moreno tied it to the sheer volume involved: “Uncovering the mysteries of the Moon requires an ability to learn from an extraordinary volume of scientific data.”
Both the model weights and the SomBench dataset are published openly, with fine-tuning code maintained in a NASA-IMPACT GitHub repository and downstream adaptation handled through TerraTorch, IBM’s geospatial fine-tuning toolkit - the same open pattern IBM and NASA have used for their earlier Earth-observation foundation models. That makes this a reusable base for outside researchers rather than a one-off internal tool: anyone building lunar mapping, landing-site assessment, or resource-prospecting software can start from a pretrained model instead of assembling and labeling their own multi-instrument lunar dataset from scratch.
The release lands as NASA and international partners are actively planning a sustained human return to the Moon, where identifying accessible ice deposits and stable terrain is a direct input to landing-site and infrastructure decisions, not an academic exercise. What it does not do is fly on a mission or make a siting decision itself; it is a research tool whose real test is whether mission teams adopt it over their existing, more manual approaches to lunar image analysis.