Multi-Modal Foundation · allenai
OlmoEarth v1.2 Small
A 36M-parameter vision transformer extracting representations from satellite imagery for Earth and climate science.
What OlmoEarth v1.2 Small is
OlmoEarth v1.2 Small is an Earth-observation foundation model developed by the Allen Institute for AI (allenai). It is an intermediate-scale model within the OlmoEarth v1.2 family, designed for tasks requiring greater capacity than the Tiny or Nano variants while requiring fewer resources than the Base model.
The model addresses remote sensing analysis across earth science, climate research, and conservation. It operates over multi-sensor satellite data, providing unified representations that researchers can compute as embeddings or adapt for downstream scientific prediction tasks.
How it works
The model employs a Vision Transformer Small (ViT-Small) architecture comprising 36 million parameters. It processes either individual satellite scenes or image time series from radar and optical sensors, specifically Sentinel-1, Sentinel-2, and Landsat.
The backbone operates as a feature extractor that can produce imagery embeddings or be fine-tuned with downstream heads for spatial prediction tasks such as semantic segmentation.
Key facts
- Architecture
- ViT-Small (Vision Transformer Small)
- Parameter count
- 36 million
- Supported sensors
- Sentinel-1, Sentinel-2, Landsat
- Input format
- Single images or image time series
- Task category
- Image feature extraction
What you can use it for
- Satellite Feature Extraction
- Generating dense embedding vectors from Sentinel-1, Sentinel-2, or Landsat imagery for clustering, similarity search, or downstream machine learning pipelines.
- Geospatial Semantic Segmentation
- Fine-tuning the model backbone to delineate land cover, habitat types, or other environmental boundaries from satellite observations.
- Multi-Temporal Satellite Analysis
- Processing satellite image time series to study dynamic environmental changes across climate and conservation studies.
Run it here
We list this model but cannot run it yet: no library and no architectures — nothing can load it.
Everything on this page still applies. Running it needs a loader for this repository, not a bigger machine.
Strengths and limits
Good at
- Provides an intermediate 36-million-parameter capacity balancing performance between smaller (Tiny and Nano) and larger (Base) variants.
- Processes data from multiple satellite platforms, including radar (Sentinel-1) and optical multispectral sensors (Sentinel-2 and Landsat).
- Accepts both single satellite scenes and multi-temporal image time series.
Watch out for
- Supported satellite inputs are restricted to Sentinel-1, Sentinel-2, and Landsat imagery.
- Serves primarily as an encoder and feature extractor, requiring task-specific heads and fine-tuning for tasks like segmentation.
- Has lower model capacity than the OlmoEarth Base model, which may affect performance on complex benchmarks.
Inputs and outputs
You provide
Image
PNG, JPEG or TIFF. Sent to the model as-is.
You get back
Mean-pooled representation, one vector per input.
Licence
Check the licence before commercial use · other
The other licence carries terms worth reading before you depend on it.
Questions
Can I use OlmoEarth v1.2 Small commercially?
Check the licence before commercial use. The other licence carries terms worth reading before you depend on it.
Which satellite sensors are compatible with OlmoEarth v1.2 Small?
The model is designed to work with imagery and image time series acquired by Sentinel-1, Sentinel-2, and Landsat.
How does the Small variant fit into the OlmoEarth v1.2 family?
OlmoEarth v1.2 Small has 36 million parameters, offering an intermediate scale for workflows requiring more capacity than the Tiny or Nano models without the resource footprint of OlmoEarth Base.
What downstream tasks can this model be fine-tuned for?
The model can be fine-tuned for remote sensing tasks such as semantic segmentation, or used directly to extract feature embeddings for classification and clustering.
How to cite it
OlmoEarth v1.2: A more efficient family of OlmoEarth models. arXiv:2605.20804.
Written from OlmoEarth v1.2 Small's own documentation (model card). Statements citing a figure that does not appear there are removed before publication. Licence and running details come from our registry.