Introducing OlmoEarth Platform: Powerful open infrastructure for planetary insights

Every second, we collect terabytes of data about our planet from satellites and sensors across the globe. These datasets hold answers to crucial questions: Where are forests disappearing? Which crops are failing? What might happen next week or next year?

Until now, those insights have remained out of reach. Outside a handful of well-resourced AI labs, planetary data rarely informs decisions. Leveraging this data requires powerful models, specialized infrastructure, and data pipelines far more complex than standard NLP or computer vision workloads. There are many powerful models, but it can be a herculean effort to make effective use of them. Doing so requires integrating disparate and often proprietary tools for data, annotation management, and orchestration—gaps that dramatically slow development and introduce barriers to adoption.

We set out to close these gaps, starting with a key insight: creating a high-performing model wasn't enough. To make Earth data useful to a broad audience, we needed to make it easy to use. So we built OlmoEarth Platform, an open, integrated, end-to-end solution that handles everything from data collection and labeling to training, inference, and deployment.

At the core of the OlmoEarth Platform is OlmoEarth, our foundation model family. OlmoEarth was pretrained on millions of Earth observations (~10 terabytes of data), delivering industry-leading performance while remaining efficient and easy to fine-tune. Unlike traditional approaches that require training separate specialized models for each task – one for crop mapping, another for deforestation, and another for land use classification, among others – OlmoEarth excels across a wide and diverse range of Earth observation applications. This multi-task versatility means organizations can use a single foundation model and customize it for multiple use cases, dramatically reducing the resources and engineering expertise needed.

Our evaluations demonstrate that OlmoEarth maintains strong performance across varied tasks without sacrificing accuracy. This capability is a fundamental advantage: Instead of managing many custom models, teams can build on one adaptable bedrock.

Just as LLMs have opened new avenues for scientific research, our goal with OlmoEarth is to unlock Earth intelligence for discovery and action. That requires openness and collaboration. OlmoEarth Platform is designed to be flexible and extensible—a cornerstone the community can build upon together. We're actively onboarding organizations that will expand what's possible, making OlmoEarth Platform and our foundation models more powerful with each new application.

OlmoEarth Studio—from data to fine-tuned models

OlmoEarth Studio – available at no cost – is your workspace for building custom Earth intelligence AI models with your extended team and partners. It supports dataset creation, fine-tuning, and insights in one place, with built-in controls to manage quality and privacy.

In Studio, you can upload your labeled data and imagery and create labeling tasks for annotators and reviewers. The platform helps you manage team members and coordinate across partners and regions, making it easier to stay aligned on large, distributed projects. 

Perhaps most valuable, Studio makes fine-tuning simple and fast. Navigate within the workspace to the model library, and – with just a few clicks – you can fine-tune our OlmoEarth foundation models, run inference, and seamlessly publish to OlmoEarth Viewer (learn more below). You also have the flexibility to configure model training settings yourself.

Behind the scenes, Studio can handle imagery acquisition from sources such as Sentinel-1, Sentinel-2, Landsat, and others. Simply specify the imagery type, area of interest, and time range, and Studio will do the rest.

Finally, Studio enables rapid iteration by letting you quickly review predictions, add labels, and retrain your model—all from a single, easy-to-use interface.

OlmoEarth Viewer—accessible public-facing artifacts

OlmoEarth Viewer is our browser-based app that enables anyone to explore model-generated maps. The goal is to create decision-ready visuals for high-value cases – such as crop mapping, land-use classification, deforestation monitoring, and ecosystem baselines – that are clear to both technical and non-technical audiences.

Here’s what you can do with OlmoEarth Viewer today:

• Compare against baseline maps using the side-by-side view, swipe bar, or transparency slider to quickly spot differences
• Scrub through time with a time slider to see how places change across dates (e.g., forest loss before versus after a fire, or seasonal crop cycles)
• Glanceable analytics show "what's in view" (class breakdowns like % cropland, % forest, and confidence levels), and update as you pan or zoom
• Publish safely from Studio through a guided review process that ensures quality and control. Draft your map, preview how it’ll appear, and then publish it with access settings that fit your needs (public, restricted link, or internal)

OlmoEarth Run—workflow engine for reproducible compute jobs

OlmoEarth Run powers the large AI training, fine-tuning, inference, and dataset building jobs that underpin the OlmoEarth Platform—reliably, at scale, and in a way you can reproduce later. 

You submit a workflow, and OlmoEarth Run partitions it into many parallel tasks, with each step declaring the code to execute, the inputs/outputs it reads and writes, and the compute it needs (for example, CPU versus GPU). 

Run automatically dispatches tasks to worker nodes, handling scheduling and resource allocation so you can target specific hardware or cloud regions without being locked into a specific dev environment. Progress and errors are tracked at the task level, with automatic retries to ensure completion.

OlmoEarth APIs—programmatic access and integration

OlmoEarth Platform provides APIs that integrate into your existing workflows, giving you programmatic control over the core steps of an Earth-intelligence project.

Upload existing datasets. Use the API to import your ground-truth data – in common geospatial formats – so you can seed annotation or train a model inside the OlmoEarth Platform.

Submit & monitor inference jobs. Once you have a custom-trained model in OlmoEarth, use the API to run inference on chosen areas of interest and to trigger jobs from events in your own systems. You can also monitor job status programmatically.

Download & integrate predictions. Retrieve model outputs via the API and pipe them into your other tools and platforms for analysis and visualization.

OlmoEarth Projects—useful fine-tuned examples

As part of the OlmoEarth release, we’re publishing tools, code, and documentation to help engineers, scientists, and researchers use our OlmoEarth foundation model family and train their own custom models. 

OlmoEarth Projects, a public GitHub repo, includes several example fine-tuned models developed with our partners, along with comprehensive guides to training models using OlmoEarth foundation models. These examples showcase OlmoEarth's breadth—demonstrating how a single pretrained model can be adapted to solve problems ranging from ecosystem monitoring to agricultural planning to disaster response. 

You can run these models entirely offline in your own environment and build new models on top of the OlmoEarth pretrained weights.

A typical project loop

Here’s what a standard project in OlmoEarth Platform looks like:

1. Import or create labels. Use annotation metadata, image attachments, CSV bulk uploads, and auto-assignment to scale labeling with built-in quality control.
2. Define & scope. Pick your area of interest and dates (e.g., western Amazon, Jan 2022-Dec 2023).
3. Fine-tune & validate. Leverage an existing recipe or create a new one, adapt a foundation model to local conditions, then compare results to available satellite imagery and iterate as needed.
4. Deploy. Publish explorable maps with access controls. Connect to serving APIs or use the download service.

Real-world impact

OlmoEarth Platform reflects a partner-led approach. By collaborating with leading organizations across forest management, wildfire resilience, food security, conservation, and ecological monitoring (to name a few), we've worked to address real pain points and streamline the path from raw data to decision-ready applications. Tools work best when they're built with the people and organizations who use them, which is why every feature reflects feedback from the field.

To that end, OlmoEarth Platform is designed for frequent data refreshes on relatively modest budgets—a major goal of many organizations. Users begin with a question and area of interest – e.g., track mangrove loss along this coast – then assemble a label set and choose an OlmoEarth model aligned to their needs. 

Partners report quicker turnarounds and costs that make regular updates feasible. Because OlmoEarth handles diverse tasks well, organizations often start with one application and expand to others using the same foundation model—maximizing their investment in resources and building institutional knowledge.

Working with NASA’s Jet Propulsion Laboratory (JPL) on wildfire risk, OlmoEarth learns from tens of thousands of field samples combined with radar and optical imagery to estimate live fuel moisture – a primary driver of ignition and spread – so planners can target prevention and response where risk is rising.

In Nandi County, Kenya, the International Food Policy Research Institute (IFPRI) adapted a crop-mapping model informed by local field data to update county-wide maps far more frequently than the historic five-year cadence, helping officials anticipate challenges, target seeds and fertilizers, dispatch field experts, and strengthen food security strategies.

For mangrove monitoring with Global Mangrove Watch (GMW), where both accuracy and update cadence are crucial, the OlmoEarth Platform has matched or exceeded industry baselines while reducing annotation requirements, enabling more frequent global updates. Near-real-time loss events enable conservationists and governments to respond faster and plan restoration where it matters most.

In the Amazon, our partnership with Amazon Conservation enables specialists to detect drivers of deforestation in near real-time, so enforcement and recovery can focus on the right places quickly.

With the Group on Earth Observations’ Global Ecosystems Atlas team, OlmoEarth helps countries in developing national biodiversity strategic action plans that incorporate up-to-date ecosystem views, enabling them to set clear priorities, allocate resources, and track progress.

Read more testimonials here and here.

Available now in early access

OlmoEarth Platform is now rolling out today.

If you’re working in food security, wildfire resilience, or on sustainability and conservation initiatives, please get in touch. We aim to empower communities to protect lives, safeguard resources, build resilience against global challenges, and thrive in harmony with nature by delivering timely, accurate intelligence.

With the OlmoEarth Platform, we’re delivering insights that are continuously updated, models that can be extended and adapted, and interfaces that anyone can use to drive impact on the ground. Join us on the journey—sign up for more information about the OlmoEarth Platform here, and explore the Viewer here.