Ai2 Ships OlmoEarth Platform and RoPE Geospatial Models

On July 28, 2026, the Allen Institute for AI (Ai2) transitioned its geospatial foundation models into an end-to-end infrastructure with the release of the OlmoEarth Platform. The platform manages the engineering lifecycle of planetary-scale Earth observation, allowing mission-driven organizations to run continent-scale inference in roughly 24 hours. By dynamically orchestrating hundreds of distributed workers to fetch multi-provider satellite pixels, the infrastructure processes dozens of terabytes of imagery at fractions of a penny per square kilometer.

Architecture and RoPE Integration

Alongside the infrastructure launch, Ai2 shipped OlmoEarth v1.2. This update directly addresses feedback from early partners by swapping absolute positional encodings (APE) for Rotary Positional Embeddings (RoPE). This architectural change eliminates the “striping artifacts” previously found in spatial representations, resulting in cleaner embeddings for high-precision mapping tasks.

The underlying foundation models utilize an encoder-decoder Vision Transformer (ViT) pretrained on 10 terabytes of multimodal data. This dataset represents millions of observations across 285,288 samples of 2.56km by 2.56km regions. Training relies on Latent MIM Lite (Latent Masked Image Modeling of Linear, Invariant Token Embeddings), which replaces the standard target encoder with a frozen, randomly initialized linear projection to stabilize the pipeline. Developers can deploy four variant sizes based on parameter counts and precision requirements: Nano (1.4M), Tiny (6.2M), Base (90M), and Large (300M).

Infrastructure and Distributed Execution

The core platform tackles the engineering bottlenecks of scaling AI inference across vast satellite data streams like Sentinel-1, Sentinel-2, Landsat, and NISAR. A new Satellite Imagery Index acts as a routing layer, maintaining pointers to cloud-optimized formats. When a developer queries for criteria like the “least-cloudy Sentinel-2 image over San Francisco,” the index returns a direct pointer to the GeoTIFF stored in Amazon S3.

For execution, the infrastructure relies on a runner Docker container that provisions virtual machines dynamically to map geographic partitions. Because distributed computing at this scale experiences routine node failures, the tasks are designed to be strictly reentrant and idempotent.

To manage this entire pipeline, the release introduces OlmoEarth Studio, a web interface designed for importing training data, managing annotations, and executing fine-tuning jobs.

Performance Gains and Deployment

Moving from standalone models to an orchestrated platform has altered the compute profile of geospatial analysis. Iterations spanning v1.1 to v1.2 reduced compute costs by up to 3x compared to the initial release, primarily driven by a 2.9x reduction in multiply-accumulate operations (MACs) for Sentinel-2 tasks.

Governments and NGOs currently deploy the infrastructure to accelerate regional mapping. The platform generates North American wildfire-risk maps, tracks cause-specific alerts for deforestation in the Amazon, and compiles country-scale crop-type maps in days rather than the months previously required.

If your organization builds mapping or remote sensing applications, evaluate the Nano and Base model variants within this new infrastructure context. The combination of RoPE-based spatial mapping and automated satellite indexing removes the need to build custom ingestion pipelines for multi-satellite imagery.