Changelog
This changelog tracks releases of the OlmoEarth pretrained foundation models. It does not cover updates to the OlmoEarth Platform.
V1.3 — October 2026
In this release, we introduce OlmoEarth v1.3, an embeddings-focused model.
- Introduced an Attention Pooler to the OlmoEarth encoder base, allowing multimodal, multitemporal tokens to be aggregated into a single per-modality embedding.
- Added a distillation head that projects the model's 768-dimensional tokens to 128 dimensions.
- Although this model focuses on embeddings, it is also a strong fine-tuning model; see our technical report for more information.
Model weights: Base
Training scripts: v1.3
V1.2 — June 2026
- Replaced fixed positional embeddings with rotary positional embeddings (RoPE), reducing artifacts in the models' embeddings.
- Modest improvement in downstream performance; read our technical report for more information.
- New Small model (35.6M parameters). Users looking for something between Tiny and Base can use Small.
Model weights: Nano · Tiny · Small · Base
Training scripts: v1.2
V1.1 — May 2026
- Up to 3× lower compute at every model size versus v1, across both training and inference, with overall performance maintained on a mix of research benchmarks.
- Collapsed multi-resolution Sentinel-2 tokens into a single token per patch (one bandset per modality instead of one per resolution), cutting sequence length roughly 3×.
- Adjusted the pretraining regimen to offset the accuracy drop that naive token merging introduced.
- Released in Base, Tiny, and Nano.
Model weights: Nano · Tiny · Base
Training scripts: v1.1
V1.0 — November 4, 2025
- Initial release of the OlmoEarth foundation models and the OlmoEarth Platform.
Model weights: Nano · Tiny · Base · Large
Training scripts: v1