Multi-Modal Foundation · allenai
OlmoEarth v1 Tiny
A 6.2M-parameter vision transformer foundation model for analysing satellite imagery and image time series.
What OlmoEarth v1 Tiny is
OlmoEarth v1 Tiny is a compact Earth-observation foundation model developed by the Allen Institute for AI. It is designed to assist researchers across earth science, climate, and conservation disciplines by providing lightweight representations of satellite observations.
The model works across data collected by Sentinel-1, Sentinel-2, and Landsat satellites, accepting either individual captures or multi-temporal image sequences. Its compact size addresses the computational challenges of executing large-scale planetary observations or running models in resource-constrained deployment environments.
How it works
The model uses a Vision Transformer Tiny (ViT-Tiny) architecture with 6.2 million parameters. It processes single images or image time series from optical and radar sensors—specifically Sentinel-1, Sentinel-2, and Landsat—to generate representations suitable for downstream feature extraction or fine-tuning.
Key facts
- Architecture
- ViT-Tiny
- Parameter count
- 6.2M
- Supported sensors
- Sentinel-1, Sentinel-2, Landsat
- Input formats
- Images and image time series
What you can use it for
- Satellite Feature Embedding
- Generate vector embeddings from Sentinel-1, Sentinel-2, or Landsat imagery and time series for downstream analysis.
- Downstream Image Segmentation
- Fine-tune the model on annotated remote sensing datasets to delineate land cover, environmental boundaries, and geographic features.
- On-Device and High-Volume Inference
- Deploy a lightweight 6.2-million-parameter backbone for efficient satellite-data processing at planetary scale or on constrained hardware.
Run it here
We list this model but cannot run it yet: no pipeline_tag — needs a human.
Everything on this page still applies. Running it needs a loader for this repository, not a bigger machine.
Strengths and limits
Good at
- Operates with a small memory footprint of 6.2 million parameters, facilitating large-scale and on-device execution.
- Supports both static individual images and multi-temporal image series.
- Processes multi-sensor inputs spanning Sentinel-1 radar as well as Sentinel-2 and Landsat optical imagery.
Watch out for
- Documented applicability is limited to Sentinel-1, Sentinel-2, and Landsat sensor inputs.
- Specialised downstream tasks such as segmentation require task-specific fine-tuning or embedding extraction workflows.
Inputs and outputs
You provide
Sequence or text
A protein or nucleotide sequence, or free text — whichever this model was trained on.
You get back
Mean-pooled representation, one vector per input.
Licence
Check the licence before commercial use · See model card
The See model card licence carries terms worth reading before you depend on it.
Questions
Can I use OlmoEarth v1 Tiny commercially?
Check the licence before commercial use. The See model card licence carries terms worth reading before you depend on it.
What satellite instruments are compatible with OlmoEarth v1 Tiny?
The model is built to handle imagery and image time series collected by Sentinel-1, Sentinel-2, and Landsat satellites.
What is the network architecture of the model?
It is based on a Vision Transformer Tiny (ViT-Tiny) architecture consisting of 6.2 million parameters.
How can researchers use the model for downstream research tasks?
Researchers can generate embeddings directly from imagery or fine-tune the architecture for remote sensing applications such as segmentation using toolkits like rslearn or olmoearth_run.
Written from OlmoEarth v1 Tiny'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.