Multi-Modal Foundation · allenai
OlmoEarth v1.2 Tiny
A compact 12-million-parameter vision model for Earth observation and remote-sensing analysis.
What OlmoEarth v1.2 Tiny is
OlmoEarth v1.2 Tiny is a lightweight Earth-observation foundation model developed by the Allen Institute for AI (Ai2). Designed for environmental and remote-sensing applications, it provides visual representations across multiple satellite imagery sources.
The model addresses the need for efficient remote-sensing feature extraction, making it suitable for large-scale analysis or on-device deployment where computational resources are constrained. It targets research and operational tasks across Earth science, climate studies, and environmental conservation.
How it works
The model is built on a Vision Transformer Tiny (ViT-Tiny) architecture comprising 12 million parameters. It processes imagery and time series collected from satellite sensors including Sentinel-1, Sentinel-2, and Landsat to extract visual feature representations.
Key facts
- Architecture
- ViT-Tiny
- Parameter count
- 12M
- Supported inputs
- Sentinel-1, Sentinel-2, and Landsat images or image time series
- Primary task
- Image feature extraction
What you can use it for
- Satellite Image Embedding Generation
- Extract compact numerical representations from Sentinel-1, Sentinel-2, or Landsat imagery for downstream machine learning pipelines.
- Land-Cover Segmentation
- Fine-tune the model representations to perform semantic segmentation across satellite scenes.
- Time-Series Remote Sensing
- Analyse temporal sequences of Earth-observation imagery to assess changes in climate and conservation contexts.
- On-Device and Large-Scale Inference
- Execute remote-sensing feature extraction in resource-constrained environments or high-throughput workflows.
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
- Possesses a very compact 12-million-parameter footprint suitable for resource-constrained or on-device inference.
- Supports multi-sensor satellite imagery across optical and radar platforms including Sentinel-1, Sentinel-2, and Landsat.
- Handles both single scenes and temporal image series for monitoring Earth-surface dynamics.
Watch out for
- May offer lower representational capacity compared to larger models in the OlmoEarth family due to its 12-million-parameter ViT-Tiny architecture.
- Requires fine-tuning or secondary models to perform specific downstream tasks such as semantic segmentation.
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 Tiny commercially?
Check the licence before commercial use. The other licence carries terms worth reading before you depend on it.
Which satellite sensors does OlmoEarth v1.2 Tiny support?
The model is designed to process imagery and temporal image series from Sentinel-1, Sentinel-2, and Landsat.
Can the model be used for image segmentation?
Yes, the foundation model can be fine-tuned directly for downstream segmentation tasks using satellite imagery.
What is the primary architectural design of the model?
It uses a Vision Transformer Tiny (ViT-Tiny) architecture containing 12 million parameters.
How to cite it
OlmoEarth v1.2: A more efficient family of OlmoEarth models, arXiv:2605.20804
Written from OlmoEarth v1.2 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.