IBM and NASA on September 10, 2026 released an open-source AI model built specifically to help scientists study the Moon, a step that space agency officials and IBM researchers say could reshape how planetary scientists search for lunar ice, map craters, and pick landing sites for future Artemis missions. The model, called the NASA-IBM Lunar Foundation Model, is available now on Hugging Face, with its full codebase posted to GitHub, according to NASA Science, IBM Newsroom, and wire coverage that followed the announcement.
The release matters beyond a single press cycle. NASA describes the system as “among the first open-source AI models built specifically for lunar science,” and both organizations say it exceeds the accuracy of widely used detection methods by up to 23% when identifying craters, potential ice deposits, and volcanic terrain on the Moon’s surface, according to IBM’s press release distributed via PR Newswire. For a space agency racing to put astronauts back on the Moon under the Artemis program, and for a company trying to prove its AI research has real-world payoffs beyond enterprise software, that 23% figure is the number worth watching.
Don't miss new tech stories on Google
Add Tech Insider once in the Google app and our stories appear in your news suggestions.
What the NASA-IBM Lunar Foundation Model actually does
At its core, the NASA-IBM Lunar Foundation Model is a geospatial foundation model, the same broad category of AI system IBM and NASA previously built for Earth observation with the Prithvi series. Instead of satellite imagery of Earth’s coastlines and forests, this model ingests decades of lunar remote-sensing data and learns general-purpose patterns that scientists can then fine-tune for narrower jobs. NASA’s announcement states the model “turns decades of lunar observations into a foundation for discovery, helping scientists surface patterns across data at a scale no single instrument has provided.”
In practice, that means three things researchers can now do faster than before. First, the model helps identify potential ice deposits inside the Moon’s permanently shadowed regions, the dark craters near the lunar poles that never see sunlight and are considered the best candidates for extractable water ice. Second, it maps craters across the lunar surface to help mission planners select safer landing sites. Third, it studies volcanic features tied to the Moon’s geological history, giving planetary scientists a clearer picture of how the lunar surface formed over billions of years, according to Engadget’s coverage of the release.
Before this tool existed, these tasks required scientists to manually sift through maps and imagery, or lean on older, lower-resolution machine-learning tools built for narrower purposes. IBM and NASA are positioning the Lunar Foundation Model as a replacement for that slower, piecemeal workflow with a single system that other researchers can fine-tune for their own specific lunar science questions, whether that’s crater cataloging, irregular mare patch analysis, or polar ice prospectivity.
The training data: 17 years of Lunar Reconnaissance Orbiter observations
The model’s foundation is roughly 17 years of data collected by NASA’s Lunar Reconnaissance Orbiter, the spacecraft that has been circling the Moon since 2009 mapping its surface in extraordinary detail. That translates to nearly 2 million image tiles in the training set, including more than 1 million 1-meter-resolution camera frames and close to 964,000 multispectral tiles, capturing everything from visible light to compositional data about the lunar surface.
NASA didn’t stop at a single mission’s data. The model draws on more than 30 layers of information collected by nine different instruments across four NASA missions, with the Lunar Reconnaissance Orbiter serving as the backbone. Supporting datasets reportedly include contributions from missions like GRAIL, which mapped the Moon’s gravity field, and Lunar Prospector, alongside data from Japan’s SELENE (Kaguya) mission run by JAXA. IBM confirmed in September 2026 that this same combination, 30-plus layers spanning nine instruments across four missions, underpins the model’s training, and that breadth is the point: a foundation model is only as useful as the diversity of signal it has seen, and stitching together nine instruments across four missions gives the Lunar Foundation Model a wider view of the Moon’s surface than any single dataset could provide on its own. It echoes the same philosophy driving growth in open source LLMs, where broader, more diverse training data tends to produce more generalizable models.
IBM and NASA also built something arguably as important as the model itself: a unified, machine-learning-ready lunar dataset. Before this project, lunar mission data existed in scattered formats across different archives, each mission’s instruments producing outputs that weren’t designed to talk to each other. IBM and NASA researchers assembled what they describe as the first open-source dataset that unifies lunar data in a form suitable for machine learning, and as of September 2026 that dataset carries its own name, SomBench, launched alongside the model as a standardized benchmark other teams can use to measure their own lunar AI systems. SomBench itself comprises roughly 2 million spatially aligned image tiles, stitched together from more than 30 datasets spanning nine instruments across four lunar missions, meaning other research teams can now train and score their own lunar-focused AI systems on a common, pre-cleaned foundation instead of repeating that data-wrangling work from scratch.
Where to find it: Hugging Face, GitHub, and the TerraTorch toolkit
Model weights and code access
NASA’s science team confirmed the model “is hosted publicly on Hugging Face for anyone to use, with the complete codebase available on GitHub for testing and experimentation.” That distribution pattern mirrors how IBM and NASA released their earlier Prithvi Earth-observation models, which have become widely used reference points in the geospatial AI research community since their debut.
Fine-tuning with TerraTorch
The Lunar Foundation Model is also integrated with TerraTorch, an open-source toolkit built for training and fine-tuning geospatial foundation models. That integration matters for adoption. A university lab or independent researcher without NASA-scale computing resources can, in theory, pull the pre-trained weights from Hugging Face, then use TerraTorch to fine-tune the model on a narrower task, such as spotting a specific type of surface feature relevant to their own research question, without having to train a lunar-scale foundation model from zero.
What isn’t public yet is the fine architectural detail some AI researchers will want: an explicit parameter count, the exact transformer variant used, or a breakdown of training compute in GPU-hours. Neither IBM’s press materials nor NASA’s science article spell out those figures, and reporting from outlets covering the release, including Reuters and Morningstar, focused on the data scale and accuracy gains rather than internal model architecture. Given the model’s lineage from IBM’s transformer-based Prithvi series, industry watchers expect a similar architectural family, but that specific detail was not confirmed in the announcement materials reviewed for this story. One open question has since been resolved: IBM and NASA released the model weights under the Apache 2.0 license, confirming the legal terms other researchers can build on even as the architectural specifics stay undisclosed. The remaining gap illustrates a broader tension in the field, one security researchers have flagged before: openly released weights don’t always come with full transparency into how a model was built.
The numbers that matter: a 23% accuracy jump
The headline performance claim from IBM and NASA is that the Lunar Foundation Model “exceeds widely used methods by up to 23% in identifying key geographic features on the Moon’s surface, including potential ice deposits, craters” and other terrain, per IBM’s official press release. Reuters’ own write-up of the September 2026 launch repeated the same figure, reporting that the model identified lunar features up to 23% more accurately than established methods, and as of October 2026 that number remains the headline statistic cited across coverage of the launch, including by financial outlets tracking IBM’s stock and strategic AI positioning.
What the 23% figure covers, and what it doesn’t
It’s worth being precise about what that number represents and what it doesn’t. IBM and NASA describe it as a comparison against “widely used methods,” which appears to reference the older, narrower machine-learning tools and manual review processes that planetary scientists have relied on until now, rather than a single named competing model. Since the September 2026 launch, that granularity has started to surface through SomBench, the unified benchmark dataset IBM released alongside the model: on ice-prospectivity mapping, the Lunar Foundation Model’s error rate fell by up to 22% compared with the SwinV2-B baseline, while crater detection accuracy improved by nearly 19% at 100-meter context resolution, according to task-level results tied to the SomBench release. Those two figures are the closest thing yet to a per-task breakdown of the headline 23% number, and further detail would likely still appear in a fuller technical paper or model card rather than the initial launch announcement.
| Metric | Detail | Source |
|---|---|---|
| Announcement date | September 10, 2026 | IBM Newsroom, NASA Science |
| Accuracy improvement | Up to 23% over widely used methods | IBM press release, Reuters |
| Primary training data source | Lunar Reconnaissance Orbiter, ~17 years of observations | NASA Science |
| Training image volume | ~2 million image tiles (1M+ camera frames, ~964,000 multispectral tiles) | Technical coverage of the release |
| Data sources | 30+ layers, 9 instruments, 4 NASA missions | Reuters |
| Hosting | Model weights on Hugging Face; code and datasets on GitHub | NASA Science, Engadget |
| Toolkit integration | TerraTorch (open-source geospatial ML toolkit) | Technical analysis of the release |
| License model | Described as open-source and publicly available | IBM, NASA |
How this connects to the Artemis program
Neither IBM’s press release nor NASA’s science article name-checks Artemis directly in the material reviewed for this story, but the stated purpose of the model lines up closely with what NASA’s human lunar exploration program needs. IBM’s own language describes the goal as helping scientists “turn decades of complex, multi-instrument data into insights to support the establishment of a sustained human presence on the Moon.” That phrase, a sustained human presence, is the long-standing shorthand NASA uses for the eventual goal beyond the initial crewed Artemis landings: a durable lunar base capable of supporting repeat missions.
Two of the model’s specific capabilities map directly onto near-term Artemis planning needs. Ice-deposit identification in permanently shadowed regions is central to in-situ resource utilization, the strategy of extracting water and other resources directly from the Moon rather than shipping everything from Earth, which would sharply cut the cost and complexity of a long-term lunar base. Crater mapping for safe landing site selection speaks directly to the operational risk NASA and its commercial partners face every time a lander approaches the lunar surface, where a boulder field or a steep crater rim invisible to a low-resolution map can end a mission before it starts.
The model doesn’t fly on a spacecraft or make real-time landing decisions. It is a research and planning tool that scientists on Earth use to analyze existing orbital data faster and more accurately than before, feeding better maps and risk assessments into the broader Artemis planning process rather than acting as an onboard autopilot.
Following the Prithvi playbook: IBM and NASA’s geospatial AI partnership
This isn’t IBM and NASA’s first open-source AI collaboration. The two organizations have spent recent years building out a family of Earth-focused geospatial foundation models under the Prithvi name, trained on satellite data for applications like climate monitoring, land-use tracking, and disaster response. Those models, also distributed openly on Hugging Face, established the template the Lunar Foundation Model now follows: take a specific type of Earth or planetary observation data, build a foundation model that captures general patterns across it, then release both the model and the underlying dataset so other researchers can build on top of it rather than starting over.
What’s new with the Lunar Foundation Model is the domain shift from Earth to another planetary body entirely. Lunar data comes with its own complications that Earth observation data doesn’t: no atmosphere to account for, permanently shadowed regions that create extreme lighting contrast, and a much smaller and more fragmented base of available imagery compared to the constant stream of Earth-observing satellites currently in orbit. Successfully adapting the geospatial foundation model approach to that different data environment is itself a notable technical achievement, even without a published architecture paper to confirm exactly how IBM’s researchers handled those differences.
Competitive and industry context: open lunar AI versus closed alternatives
The decision to release the Lunar Foundation Model openly, rather than keep it as an internal NASA research tool or a proprietary IBM product, sets it apart from how commercial AI companies typically handle specialized scientific models. Most large AI labs building domain-specific systems, whether for drug discovery, materials science, or geospatial analysis, tend to keep the highest-performing versions behind an API or an enterprise license, a pattern visible in releases like Tencent’s Hy3 coding model and other open-weight systems that still carry usage restrictions. IBM and NASA chose the opposite path for lunar science, publishing both the model weights and the training dataset for free use by any researcher, university, or space agency worldwide.
That openness is strategically useful for both parties. For NASA, an open model invites contributions and improvements from the global planetary science community, potentially surfacing insights the agency’s own researchers wouldn’t have found working alone, and it aligns with the agency’s public mandate to share the results of taxpayer-funded science. For IBM, positioning itself as the AI partner behind foundational, publicly available scientific models builds a research reputation distinct from the consumer chatbot and enterprise-copilot race that dominates coverage of companies like OpenAI, Google, and Anthropic. IBM doesn’t compete directly in that consumer AI assistant market, so headline scientific applications like a lunar foundation model serve as a different kind of proof point for its AI research division.
| Approach | Example | Access model | Primary use case |
|---|---|---|---|
| Open lunar geospatial model | NASA-IBM Lunar Foundation Model | Free, Hugging Face + GitHub | Crater mapping, ice detection, terrain analysis |
| Open Earth geospatial model | IBM-NASA Prithvi series | Free, Hugging Face | Climate, land-use, disaster response |
| General-purpose frontier AI | Large commercial chat/agent models | Paid API or subscription | General reasoning, coding, enterprise tasks |
| Proprietary scientific AI tools | In-house lab or agency research systems | Internal only, not published | Narrow, mission-specific research questions |
What planetary scientists and industry analysts are saying
IBM’s newsroom framed the release in broad terms, stating that “IBM and NASA today announced the open-source release of the NASA-IBM Lunar Foundation Model, one of the first publicly available foundation models for scientific exploration of the Moon, now available,” according to IBM Newsroom. NASA’s own science communications team echoed that framing, describing the system as “among the first open-source AI models built specifically for lunar science,” per NASA Science.
NASA also detailed the data lineage behind the release, noting that “the model, trained primarily on data from NASA’s Lunar Reconnaissance Orbiter (LRO), is hosted publicly on Hugging Face for anyone to use, with the complete codebase available on GitHub for testing and experimentation,” according to NASA Science.
Wire coverage focused on the practical research payoff. Reuters described the tool plainly: “The NASA-IBM Lunar Foundation Model is a publicly available AI tool designed to study the Moon,” and detailed how “it can help researchers identify potential ice deposits on the Moon’s permanently shadowed regions, map craters to select safe landing sites and study volcanic features, tasks that have traditionally required scientists to sift through maps and images manually or rely on lower-resolution machine-learning tools,” according to Reuters. That framing, replacing slow manual review with a faster automated first pass, is the core value proposition industry watchers have pointed to since the launch.
The IBM stock and business angle
Financial press picked up the announcement quickly, with Morningstar covering it as part of IBM’s broader AI strategy positioning rather than purely a science story. That framing reflects a real dynamic: IBM has spent the past several years working to convince investors and enterprise customers that its AI research translates into deployable, differentiated products, at a time when the company’s public AI narrative is often overshadowed by the trillion-dollar spending figures coming out of Nvidia, Microsoft, Google, and OpenAI. A high-profile, scientifically credible open-source release with a concrete performance number attached, the 23% accuracy improvement, gives IBM’s communications team a tangible proof point that doesn’t require comparing IBM’s consumer-facing AI products directly against ChatGPT or Gemini, a fight it has largely opted out of. It’s a different strategic bet than the one Nvidia made with its reported Hugging Face acquisition, which centers on owning distribution rather than publishing standalone scientific models.
For NASA, the calculus is different but complementary. The agency operates under sustained budget scrutiny, and demonstrating that its decades of Lunar Reconnaissance Orbiter data are being put to new, high-value use through a modern AI partnership helps justify continued investment in both the orbiter program and future data-collection missions. An open-source release that other researchers worldwide can immediately use and build on is also a low-cost way for NASA to multiply the scientific return on data it already spent public money to collect.
Historical context: from Apollo-era maps to AI-assisted lunar science
It’s worth stepping back to appreciate how far lunar mapping has come. During the Apollo program in the 1960s and 1970s, mission planners worked from photographic mosaics and hand-drawn geological maps, with landing site selection driven by a relatively small number of high-resolution images and a lot of expert judgment under time pressure. The Lunar Reconnaissance Orbiter, launched in 2009, changed that entirely by providing a continuous, high-resolution stream of imagery and instrument data covering nearly the entire lunar surface over more than a decade and a half of operation.
The bottleneck shifted from data collection to data analysis. With millions of image tiles and multiple instrument types generating information faster than human researchers could manually review it, the case for machine learning became obvious well before this specific model existed. IBM and NASA’s earlier work on Prithvi for Earth observation demonstrated that a foundation-model approach could meaningfully speed up analysis of large remote-sensing datasets. Applying that same architecture and philosophy to 17 years of accumulated lunar data represents the natural next step in that progression, arriving just as NASA’s renewed human lunar exploration ambitions under Artemis make fast, accurate lunar terrain analysis newly urgent.
What researchers can build on top of it
Because the Lunar Foundation Model is released as a foundation model rather than a finished single-purpose application, its real value will likely show up in the coming months as independent research teams fine-tune it for specialized tasks. NASA’s own framing points toward several directions: crater mapping refinements, analysis of irregular mare patches (unusual smooth, young-looking volcanic deposits scattered across the lunar surface whose origin remains debated), and polar ice prospectivity mapping to help narrow down the best sites for future resource-extraction missions.
University planetary science departments, international space agencies without their own large-scale AI research budgets, and even student researchers now have access to a pre-trained lunar-specific model they can adapt with comparatively modest computing resources through TerraTorch, rather than needing to train a foundation model from scratch on raw Lunar Reconnaissance Orbiter data, a similar lowering of the barrier to entry that open-weight releases like Meta’s Muse Glimmer have brought to more general-purpose AI development. That democratization effect, lowering the resource bar for doing sophisticated lunar data analysis, may end up being a bigger long-term impact than the headline 23% accuracy figure from the initial announcement.
Predictions: where this goes from here
Based on the pattern set by IBM and NASA’s earlier Prithvi releases and the stated goals behind this launch, a few developments look likely in the months ahead. First, expect fine-tuned derivative models to appear on Hugging Face within the next few months, built by university labs or international partners applying the base Lunar Foundation Model to narrower tasks like specific crater classification or targeted ice-stability mapping in particular polar regions.
Second, expect NASA to reference outputs from the model, or from derivative fine-tuned versions, in future public communications about Artemis landing site selection, particularly for missions targeting the lunar south pole where permanently shadowed regions and ice prospecting are most relevant. Third, given IBM’s pattern with Prithvi, a peer-reviewed technical paper detailing the model’s architecture, exact parameter count, and full benchmark results will likely follow the initial press announcement, filling in the technical gaps that current coverage hasn’t addressed. Fourth, other national space agencies with active lunar programs may look to either use the open model directly or adopt a similar open-foundation-model approach for their own planetary datasets, following the credibility IBM and NASA have built with this release. Fifth, expect this release to feature in IBM’s ongoing case for its AI research credibility in future enterprise sales and investor conversations, given how directly it ties IBM’s brand to a scientifically significant, non-consumer AI application.
Why open-source matters for planetary science specifically
Planetary science operates differently from many other AI application domains in one important way: the underlying data, imagery and measurements of the Moon, Mars, and other bodies, is collected by a small number of government-funded missions and is largely non-commercial. There’s no natural market incentive for a private company to build a specialized lunar AI model the way there is for, say, a coding assistant or an image generator with obvious consumer demand. That makes government-industry partnerships like this one between NASA and IBM one of the few realistic paths to getting sophisticated, purpose-built AI tools into this scientific domain at all.
Releasing the model and dataset openly also sidesteps a common criticism leveled at AI research generally: that impressive capabilities get locked behind corporate paywalls, limiting who can actually benefit from the research. A planetary scientist at a smaller international university, without access to a well-funded AI lab’s proprietary tools, can now pull the same NASA-IBM Lunar Foundation Model that a NASA-affiliated researcher would use, a meaningfully different dynamic than what exists in most corners of the AI industry today.
Remaining questions and gaps in public information
Several technical details remain unconfirmed in the public materials released alongside the model. IBM and NASA have not published an explicit parameter count, a specified transformer architecture variant, or a breakdown of the training compute used to build the model. The licensing question, at least, is now settled: IBM and NASA released the model weights under the Apache 2.0 license, confirming that both organizations are treating the release as genuinely open-source and freely usable rather than open-weight with restrictive terms attached.
The precise benchmark methodology behind the 23% accuracy improvement figure is now partially broken down by task through SomBench, the lunar benchmark dataset IBM launched in September 2026, which credits the model with up to a 22% reduction in ice-prospectivity error against the SwinV2-B baseline and close to a 19% gain in crater detection at 100-meter context resolution. A fuller technical paper or model card, which, based on IBM and NASA’s precedent with the Prithvi series, typically follows an initial press announcement like this one within a matter of months, would still help confirm the full methodology behind those numbers.
Frequently asked questions
What is the NASA-IBM Lunar Foundation Model?
It’s an open-source AI geospatial foundation model built by IBM and NASA, trained on decades of lunar observation data, designed to help scientists identify craters, potential ice deposits, and volcanic terrain on the Moon’s surface.
When was the NASA-IBM Lunar Foundation Model released?
IBM and NASA announced the open-source release on September 10, 2026, according to IBM Newsroom and NASA Science.
Where can I download the NASA-IBM Lunar Foundation Model?
The model weights are hosted publicly on Hugging Face, and the full codebase is available on GitHub, according to NASA’s science communications team.
What data was the model trained on?
Primarily around 17 years of data from NASA’s Lunar Reconnaissance Orbiter, roughly 2 million image tiles including more than 1 million 1-meter camera frames and nearly 964,000 multispectral tiles, supplemented by data from other NASA missions and instruments.
How accurate is the model compared to previous methods?
IBM and NASA state it exceeds widely used methods by up to 23% in identifying key geographic features on the lunar surface, including ice deposits, craters, and volcanic terrain.
Is this related to IBM’s Prithvi AI models?
Yes. The Lunar Foundation Model extends the geospatial foundation model approach IBM and NASA previously used for Earth observation with the Prithvi series, applied here to lunar data instead.
Does this model help NASA’s Artemis program?
The model’s stated purpose, supporting a sustained human presence on the Moon through better ice-deposit identification and crater mapping for landing site selection, directly supports the kind of planning work that underpins NASA’s broader lunar exploration goals.
Can independent researchers fine-tune the model for their own projects?
Yes. The model is integrated with the open-source TerraTorch toolkit, allowing researchers to fine-tune the pre-trained model for specialized tasks like crater cataloging or polar ice prospectivity without training a new foundation model from scratch.
