How to fine-tune SatlasPretrain on Sentinel-2

How it is run today: Self-host open weights or process code

SatlasPretrain (Ai2 / Allen Institute) publishes pretrained backbones for satellite and aerial imagery, including Sentinel-2, to fine-tune for land cover, crops, buildings, and similar tasks.

There is no hosted Satlas inference API that we found. Fine-tuning Sentinel-2 is self-host work: take a backbone, add a customer head, bring labels.

To deliver: host the backbone plus a customer head. GPU class is unknown. Model weights are ODC-BY (commercial use allowed with attribution; verify ODC-BY obligations). The code package is Apache-2.0. The customer supplies imagery matching a supported modality, plus labels for the fine-tune.

We can stand that up after a named job. We do not already host Satlas. We did not train the backbone. We do not guarantee a land-cover map.

Upstream

Also in the catalog : SatlasPretrain (Ai2).

Reach out

If this is the job you need, say so. We will not claim we already host it. We will say whether we pass through, stand it up, or cannot do it honestly.