Multi-Modal Foundation · allenai
OlmoEarth v1.2 Nano
A compact remote-sensing foundation model extracting features from Sentinel and Landsat satellite imagery.
What OlmoEarth v1.2 Nano is
OlmoEarth v1.2 Nano is an Earth-observation foundation model developed by the Allen Institute for AI. Designed as a minimal-footprint checkpoint, it targets high-throughput inference pipelines and embedded computing across earth science, climate monitoring, and conservation research.
The model handles remote sensing data acquired by Sentinel-1, Sentinel-2, and Landsat platforms, processing both individual static scenes and multi-temporal image series. Researchers can use it directly to produce general-purpose image representations or fine-tune it for specific downstream spatial tasks such as segmentation.
How it works
The model is built on a ViT-Nano (Vision Transformer Nano) architecture comprising 1.7 million parameters. It extracts image features from multi-modal and multi-temporal satellite inputs, including synthetic aperture radar from Sentinel-1 and multispectral optical imagery from Sentinel-2 and Landsat.
Key facts
- Architecture
- ViT-Nano
- Parameter count
- 1.7M
- Supported inputs
- Sentinel-1, Sentinel-2, and Landsat imagery or image time series
- Primary task
- Image feature extraction
What you can use it for
- Satellite Feature Extraction
- Generating compact numerical embeddings from Sentinel-1, Sentinel-2, and Landsat acquisitions for downstream statistical analysis or clustering.
- Image Segmentation
- Fine-tuning the backbone on task-specific labels to segment terrestrial structures, land cover, or environmental changes.
- High-Throughput Remote Sensing
- Executing rapid inference across vast archives of satellite imagery or on hardware-constrained edge and embedded systems.
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
- Minimal 1.7-million parameter footprint enables high-throughput processing and embedded deployment.
- Native support for multiple satellite systems, including Sentinel-1 radar and Sentinel-2 and Landsat optical sensors.
- Capable of processing both single static acquisitions and temporal image sequences.
Watch out for
- Focuses on feature extraction and requires fine-tuning or secondary task heads for concrete applications like segmentation.
- Restricted to inputs originating from Sentinel-1, Sentinel-2, and Landsat platforms.
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 Nano 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 this model?
The model accepts image and time series inputs from Sentinel-1, Sentinel-2, and Landsat platforms.
What architecture and size is the model?
It uses a Vision Transformer Nano (ViT-Nano) architecture containing 1.7 million parameters.
Can the model be adapted to custom tasks?
Yes, it can be fine-tuned for downstream remote sensing objectives, such as semantic segmentation, or used directly to generate feature embeddings.
How to cite it
OlmoEarth v1.2: A more efficient family of OlmoEarth models, arXiv:2605.20804
Written from OlmoEarth v1.2 Nano'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.