​
Clay model

Clay model is a 200 million parameter open-source transformer generating embedding representation for any region on the planet. It was trained on a variety of open-source satellite data with a tile size of 256x256. You can read more about it on Clay model developer webpage.

​
Embedding Inference

Our embedding inference API returns a 768-dimensional vector representation for any given satellite image. This embedding is useful for many various downstream tasks. We provide access to raw embeddings endpoints for your own experimentatio. Each API call, whether using raw pixels or coordinates, returns normalized 768-vector output.

Raw Pixel Inference

Pass raw satellite pixels (must be sqare with the length divisble by 8) and metadata directly for inference. Ideal for users with their own satellite imagery.

Coordinate-based Inference

Provide latitude and longitude coordinates. We automatically retrieve satellite data and perform inference.

Text Inference

Provide text and we will retrieve an embedding for it via SkyCLIP. You can use it for similarity comparisons with image embeddings.

​
Finetuning for downstream tasks

Our fine-tuning API empowers you to customize the Clay model for specific geospatial tasks, supporting regression, classification, and segmentation workflows. Simply provide your labeled dataset and configuration, and we’ll handle the fine-tuning process. Once complete, we offer hosted inference for your fine-tuned models, enabling fast and efficient predictions via raw pixel input for your custom tasks.

Classification

Develop models for categorizing land use, detecting specific structures, or identifying environmental features.

Segmentation

Create pixel-level classification models for tasks like urban area delineation, forest cover mapping, or crop field boundary detection.