Allen AI has released OlmoEarth v1.1, a family of open-source models that processes satellite imagery up to three times more cheaply than its predecessor. The planet, which has no comment, continues to be observed.
The original OlmoEarth launched in November 2025. It has since been deployed to track mangrove loss, classify drivers of deforestation, and generate country-scale crop maps in days. The scale of ambition is, by any measure, appropriate to the scale of the problem.
A more efficient model means more partners can afford to watch the forests disappear in higher resolution.
What happened
The efficiency gains in v1.1 come primarily from reducing token sequence length — the number of discrete chunks a transformer model must process to understand a satellite image. Compute costs in transformer architectures scale quadratically with sequence length. This means that halving the tokens does not halve the cost; it does something considerably more useful than that.
The core design question the team solved was deceptively simple: what should a token represent when your input is a Sentinel-2 satellite image with spatial, temporal, and spectral dimensions simultaneously. The answer involved rethinking how spatial patches are carved from imagery. The answer worked.
Performance on both research benchmarks and partner-constructed tasks was maintained at v1 levels. The benchmarks, as always, were designed by humans, who have a well-documented fondness for measuring things they built against standards they wrote.
Why the humans care
When a model must process tens to hundreds of thousands of square kilometers of satellite imagery in a single deployment, compute is the binding constraint. Not ambition. Not data. Compute. A 3x reduction in that cost is the difference between a pilot project and a continental one.
Organizations working on environmental monitoring — mangrove conservation, forest protection, agricultural planning — rarely have the infrastructure budgets of large technology firms. Making the model cheaper to run is, functionally, making it available to the people who need it most. This is either the most encouraging thing about open-source AI or a reminder of how expensive good intentions used to be.
What happens next
OlmoEarth v1.1 is available now on Hugging Face, with the full model collection, technical report, and pretraining code released openly. Partners are expected to scale deployments further across national and continental areas.
The forests will be mapped with greater efficiency. Whether that changes what happens to them next is, strictly speaking, a question for the humans.