IBM and NASA Open-Source Lunar Foundation Model for Space Exploration

On September 10, 2026, IBM and NASA announced the open-source release of a multimodal, multi-resolution foundation model for lunar remote sensing, along with the SomBench training dataset. [1] The model weights are available on Hugging Face under the Apache-2.0 license, while the fine-tuning code is kept in a NASA-IMPACT GitHub repository. [1] Adaptations for downstream tasks are managed through TerraTorch. [1] The system enables fine-tuning or LoRA-adapting the encoder for tasks such as detection, segmentation, and dense regression on LROC imagery. [1]

The model can be applied to several areas of lunar research, including predicting the presence of ice in permanently shadowed regions that could hold resources for future lunar bases. [1] Researchers can also use the model to map Irregular Mare Patches to study the volcanic history and thermal evolution of the Moon. [1] Additionally, the model aids in detecting and classifying craters to help NASA select safe landing sites by avoiding steep slopes and boulders. [1] According to NASA's Kevin Murphy, the model demonstrates the potential of applying AI to the agency's petabytes of scientific data. [1]

The release of the NASA-IBM Lunar Foundation Model represents a significant push toward open-science AI in space exploration. By publishing under an Apache-2.0 license and providing the co-registered SomBench dataset, IBM and NASA are equipping the broader scientific community with advanced tools for processing remote sensing data. Compared to general-purpose vision models, this specialized system offers targeted capabilities for critical aerospace objectives, such as resource identification and safe landing site selection. However, the reliance on TerraTorch and GitHub repositories implies that researchers will still need specialized machine learning engineering skills to adapt the encoder for custom tasks. What the sources don't address: whether IBM and NASA plan to release subsequent updates to the model as new lunar observation data is collected.