OlmoEarth v1.1 cuts AI costs by up to 3x
Ai2 has introduced OlmoEarth v1.1, a family of models for analyzing satellite imagery with up to three times less computation than the previous version. It maintains similar performance across several benchmarks, though it may show regressions on certain tasks.
Allen Institute for AI (Ai2) has introduced OlmoEarth v1.1, a new family of models that analyzes satellite imagery with up to three times less computation than the previous version, according to its creators.
The tool is designed for tasks such as tracking mangrove changes, identifying the causes of forest loss, and creating crop maps at national scale. Its first version, released in November 2025, had already been used in projects spanning national, continental, and global scales.
The goal of this update is practical: make those analyses faster and cheaper. When a model processes images covering tens or hundreds of thousands of square kilometers, the biggest cost often comes from the computation needed to prepare the data, run the model, and process the results.
Less repeated data, lower costs
OlmoEarth uses an architecture called a transformer, which is common in modern AI systems. To work with satellite imagery, it first converts the data into small units called tokens that the model can process.
In images from Sentinel-2, one of the most widely used satellites for observing Earth's surface, the previous version generated separate tokens for each spatial resolution. Because Sentinel-2 offers three main resolutions, the same area could produce three tokens for each observation time.
OlmoEarth v1.1 combines those resolutions into a single token. The result is a sequence up to three times shorter. This matters because transformer costs grow quadratically as sequence length increases: reducing the number of tokens does not just eliminate work, it can reduce it considerably.
The change was not as simple as combining the data. Early tests caused a 10 percentage point drop on m-Eurosat kNN, a common benchmark for evaluating Earth observation models. Ai2 modified the training process to recover the lost performance.
What changes for people using the model
The new family maintains performance similar to OlmoEarth v1 across a combination of research benchmarks and tasks created with partner organizations. The main promise is that each model size can run at up to one-third of the previous computational cost.
That could mean:
- Updating forest or crop maps more frequently.
- Processing large territories without waiting as long.
- Reducing the cost of fine-tuning and running the model.
- Making geospatial AI more accessible to teams with less infrastructure.
Ai2 warns that some regressions do exist on certain tasks. For that reason, organizations already using the original family should check whether OlmoEarth v1.1 performs equally well on their own data before migrating.
The update includes Base, Tiny, and Nano models, along with their pretrained weights and the code needed to train them. The version also uses the same dataset as OlmoEarth v1, allowing researchers to better study how much of the change comes from the architecture and how much comes from the training method.
The underlying idea is simple: using AI to observe the planet depends not only on whether the model gets the answer right, but also on whether it can run at a useful scale. If OlmoEarth v1.1 maintains its accuracy across more use cases, the next step will be to update global maps more frequently at a cost more organizations can afford.