Ai2’s OlmoEarth Opens the Door to Accessible, Open-Source Earth AI

The Allen Institute for AI (Ai2) is announcing the release of OlmoEarth, a new platform featuring an open multimodal foundation model pre-trained on approximately 10 terabytes of global satellite and sensor data. Developers can utilize this AI to gain a better understanding of what’s happening to our planet, as it democratizes access to insights that were traditionally out of reach, mostly available only to well-resourced labs and institutions.

OlmoEarth is more than its eponymous large language model. The platform also includes a studio for creating custom Earth intelligence models, a browser-based app that provides an outlet for exploring model-generated maps, and a workflow engine for scheduling reproducible compute jobs. Ai2 is also releasing APIs that developers can use to integrate OlmoEarth into their own tooling natively.

“The challenges facing our planet today demand solutions that are not only powerful but are accessible to all,” Ali Farhadi, Ai2’s chief executive, remarks in a release. “With OlmoEarth, we’re making Earth AI accessible to those working on the front lines. It’s built to help governments, [non-governmental organizations], and communities that wouldn’t otherwise have access to AI. It enables them to use it through an adaptable and open platform—allowing them to see change as it happens and empowering them to respond quickly in a time of need.”

With this platform, Ai2 continues its push to use AI to advance conversation work, one of the nonprofit’s three main focuses. Along with OlmoEarth, the lab has developed EarthRanger, its wildlife conservation software, and Skylight, which tracks illegal fishing and safeguards marine ecosystems. Notably, Joseph Redmon, an Ai2 research scientist, reveals in an interview that OlmoEarth was born out of the lab’s Earth Systems project, which launched in 2024 as a multimodal geospatial intelligence platform.

As he tells me, Ai2 has been assisting non-profit organizations by producing special-purpose models for their use. “It’s not very scalable if we’re the ones doing all of the training behind it,” he concedes. “Part of the idea for OlmoEarth was, can we figure out what the core components are of this pipeline and make a platform that can not make us do all this work, but allow these partner organizations to do the work themselves, to enable these organizations that…have such high levels of expertise in their domains?”

“They have their own domain knowledge and their expertise, and we can build this platform that they can bring all that knowledge and expertise to and actually be able to use all of these advances that we’ve made in deep learning, machine learning, in training these big models, Redmon goes on to add. “So the idea came like, we have all these organizations that want to work with us. We don’t have the time to work with them all, but we can build this tool that they can do everything that we can do, themselves, basically.”

The OlmoEarth Model Family

Developers can choose from four versions of OlmoEarth: OlmoEarth-v1-Nano (approximately 1.4 million parameters), OlmoEarth-v1-Tiny (approximately 6.2 million parameters), OlmoEarth-v1-Base (90 million parameters), and OlmoEarth-v1-Large (300 million parameters)—all built on the Vision Transformer (ViT) architecture, which was specifically designed for computer vision tasks. These are only the beginning, as Ai2 reveals it’s already working on next-generation OlmoEarth models that it anticipates will be released next year and will expand into new sectors, such as humanitarian response, with support for weather data and other modalities.

Redmon states that the four variations were chosen based on “previous literature” about the types and sizes of models that perform well in image processing, each balancing speed, accuracy, and efficiency differently. OlmoEarth-v1-Nano, for instance, offers the fastest and most efficient inference, while OlmoEarth-v1-Large is the most computationally expensive, but typically delivers the best results.

Ai2 claims that OlmoEarth outperforms larger commercial and academic models, such as Meta’s DINOv3, IBM/NASA’s Prithvi, IBM’s Terramind, CROMA, and Panopticon, as well as its previously released geospatial models, Satlas and Galileo. OlmoEarth excels in key tasks, including crop mapping, mangrove extent estimation, and live fuel moisture predictions. All of these are considered valuable because they assess a model’s ability to generalize real-world geospatial data that directly affects climate, ecosystems, and human livelihoods—all critical elements when thinking about the planet.

Redmon adds that OlmoEarth’s large model “has the best performance on a majority of tasks,” while the base model “is pretty good on a lot of stuff,” and the tiny version “sometimes has some good performance.”

Building OlmoEarth wasn’t without its challenges. In fact, although it uses ViT, planetary data isn’t necessarily compatible with the image files that computers may typically analyze.

“These transformers were designed for natural image tasks, and Earth observation is actually fairly significantly different from natural images, so they might not be the precisely correct sizes,” explains Redmon. “We think there’s probably a lot of efficiency gains that we can make, investigating the actual structures of these.”

Unlike traditional language data that goes into models like ChatGPT or Claude, Earth observation data has complex multi-dimensional characteristics, including temporal and spatial components, where data shows changes over time and geographical positioning; and multiple modalities, pulling from different satellite data sources (e.g., Sentinel-1, Sentinel-2, and Landsat) that are refreshed every seven days. In other words, while language models learn patterns in words and images, Earth AI models must understand patterns in the planet itself, tracking how landscapes evolve, ecosystems shift, and climate signals emerge over time.

OlmoEarth was trained on 10 terabytes of publicly available satellite data—much larger than most text corpora used for language models, yet still relatively modest for planetary-scale datasets. It’s not something that must run on cloud infrastructure, so don’t expect OlmoEarth to be made for edge devices. Still, why did Ai2 choose 10 terabytes as its sample size? Simply, it found that “it wasn’t that useful to scale that much higher.”

“Satellite data is, like a lot of distributions, has a really long tail. But the bulk of the data—if you’ve seen one ocean tile, you’ve seen all of the ocean tiles. If you’ve seen one forest, you’ve seen at least a lot of the forest that is in that local area. So there’s a lot of redundancy in satellite data. We sampled a very large dataset and trained on it,” Redmon details. “If we wanted to get improvements from adding more data, it’s not as simple as turning the knob up. There’s so much data out there that we could easily get more data for our training set, but we’ve seen that it doesn’t improve performance significantly without changing how we’re sampling. We probably could get increased performance if we had the ability to sample basically from that long tail and find some of these more interesting and rare pieces of what is kind of out there. But at this point, we have found a training set that seems to work very well for us.”

Building, Viewing, and Automating Planetary Data

Besides the OlmoEarth model family, there are four additional aspects of this platform, which Redmon refers to as “data labeling, fine-tuning, running inference, and viewing the results.” All are designed to help solve an Earth observation task.

OlmoEarth Studio

The experience begins with OlmoEarth Studio, a workspace platform that allows users to customize their own models. They can upload proprietary labeled data and imagery to fine-tune their chosen OlmoEarth model, as well as set up tasks for annotators and reviewers.

“The idea is, you either bring your data, or [if] you don’t have any data and you just bring the expertise, and you can label the data within the platform,” Redmon says. “Either way, the data gets in there, and then you’re fine-tuning your model on top of our foundation models, which we’ve already pre-trained, which we know have high-performance across a wide variety of different tasks.”

As is typical with AI tools today, OlmoEarth Studio requires little coding knowledge. Ai2 shares that with a few clicks, users can craft their own custom AI, run predictions, handle annotations, and then easily publish it to OlmoEarth Viewer.

OlmoEarth Viewer

That leads to the next part of this geospatial intelligence platform: a browser-based app that enables anyone to explore model-generated maps. It’s one thing to modify a ViT model that’s made up of file names and data, but what does the output look like? That’s where OlmoEarth Viewer comes into play—think of it like the Google Earth interface.

According to Ai2, the app allows users to compare new imagery with baseline maps, travel through time to see how environments have transformed—such as forests before and after wildfires—and produce analytics showing what’s in view, including the proportion of cropland or forest, which refreshes when the map is moved. Users control the map they want generated, including its appearance and how it’s published (e.g., public, restricted to those with a link, or for internal use only).

OlmoEarth Run

OlmoEarth Run is the platform’s automation builder, turning developer workflows into a series of parallel tasks. Each step includes its own executable code, input and output specifications, and compute requirements. OlmoEarth Run acts as an orchestrator, supervising scheduling and resource allocation, while tracking progress and errors at the task level.

“It’s the interface, basically, for saying, I have this model. I’ve already fine-tuned it on this data. I want to run it on this section of the world, and I want it to run over these particular time steps,” Redmon explains. “It’s our basic tool and interface for taking models and either fine-tuning them on data or running inference on some particular part of the world.”

Ai2 discloses that this tool is still in development, but it does currently support “some configurations of automated fine-tuning.” Additional capabilities will be added once more organizations are onboarded.

OlmoEarth API

No developer platform would be complete without extensibility. OlmoEarth includes built-in APIs that let organizations integrate it into existing workflows and programmatically control its core functions. When connected, users can upload existing datasets to have OlmoEarth’s ViT models grounded on their internal data. They can also utilize the API to perform inference on specific areas of interest or trigger jobs based on events. Lastly, users can retrieve model outputs and pipe them into other tools and platforms for analysis and visualizations.

Who’s Using OlmoEarth?

Ai2 shares that its OlmoEarth platform is designed for organizations that require frequent geospatial data refreshes but may not have the extensive budget to support them.

For example, the International Food Policy Research Institute in Kenya is using OlmoEarth to develop a crop-mapping model trained on local field data, providing county-wide maps issued far more frequently than the historic five-year cadence. Officials could use them to proactively address challenges, target seeds and fertilizers, and strengthen food security strategies.

The Global Mangrove Watch organization is also using OlmoEarth to monitor the growth of mangroves, an endangered ecosystem. Ai2 claims its platform has “matched or exceeded industry baselines while reducing annotation requirements.” The accuracy and rapid cadence are enabling conservationists and governments to react more quickly and plan restoration in key areas.

That said, OlmoEarth isn’t the only Earth intelligence offering on the market today. What advantage does an open-source platform have over incumbents? One “obvious” rival highlighted by Redmon is the AlphaEarth Foundations, an AI model developed by Google DeepMind. He classifies it as a closed model with “no plans to ever release the trained version,” and any embeddings provided are on a one-year timeframe that comes with a cost. When Ai2 compares OlmoEarth against AlphaEarth, Redmon boasts that both are on par with each other. However, the edge OlmoEarth has is that “you can fine-tune our model.”

“In terms of [the] advantage over proprietary systems, just being able to fine-tune the model is a big one, and an obvious one,” he points out. “Of course, people want to do this, and maybe at some point [Google DeepMind] will allow users to do sort of similar things that fine-tune their model. It would be great if they open-source their model. It’d be interesting. But beyond that, fine-tuning makes these models better.”

Redmon also claims that open-source builds better trust with organizations, eliminating the fear of what could happen if a vendor ends support or goes out of business. “One of the benefits is that we can go to these partners and be like, ‘All of the stuff that we’re doing is open-source. We are here now, and we want to support you and help you run and train these models. But, if something happens and Ai2 is gone next year for some reason, all of this stuff is still there—you can keep using it. You have the models that you’ve trained. You can keep running them on your data.’ We’re trying to make it as easy as possible to take these models and run them on your own cloud infrastructure if you want, or if you have a desktop with a GPU in it.”

The OlmoEarth platform is now available. Ai2 is releasing its code, documentation, and example models through its public GitHub repository, OlmoEarth Projects.

Featured Image: Ai2 has launched OlmoEarth, an open platform using AI and satellite data to provide planetary intelligence. Credit: Ai2