AllenAI releases OlmoEarth v1.1, a family of open transformer models for Sentinel-2 satellite imagery analysis that maintains state-of-the-art performance while cutting compute costs by up to 3x

OlmoEarth v1.1 just dropped (thx @allen_ai) 🌍

This family of Earth observation foundation models for satellite imagery tasks (e.g. mangrove change tracking, forest loss driver classification) just got 3X CHEAPER/FASTER to run.

The trick is redesigning what a token represents. Sentinel-2 inputs used to get one token per resolution (10m/20m/60m). v1.1 collapses them → 3x fewer tokens, quadratically cheaper compute.

Available now in the same sizes as v1: Nano, Tiny, Base. Open weights, open training code.

If you're running v1 and v1.1 works for your task, expect significant speedups during fine-tuning & inference.

🤗 Models: https://huggingface.co/collections/allenai/olmoearth 🔗 Blog: https://allenai.org/blog/olmoearth-v1-1

Against the constant pressure of *genAI, genAI, genAI*, I am really appreciating @allen_ai 's work on creating tools for critical needs -- like crop maps and forest loss analysis. They just did a nice release on @huggingface , check it out (linked below)

Compute is the highest cost when running OlmoEarth at hundreds of thousands of square kilometers. Partners use v1 today for mangrove tracking, forest-loss classification, & country-scale crop-type mapping.

v1.1 makes that work more sustainable.

Done naively, this hurts accuracy noticeably. Recovering it took changes to how we pretrain the model.

→ Full details in our tech report: https://allenai.org/papers/olmoearth_v1_1

Plus, they made a great blog if you want to dive a bit deeper

https://huggingface.co/blog/allenai/olmoearth-v1-1

A useful property for researchers: we held the pretraining dataset constant from v1.

The differences cleanly isolate the methodological change, not the data or the architecture family.

https://huggingface.co/blog/allenai/olmoearth-v1-1

Where the savings come from: we feed the model about 3x fewer tokens per Sentinel-2 input.

Since compute scales quadratically with token count, even modest reductions compound into real efficiency gains.

@allen_ai This is as always incredible work! I could not locate the dataset, is that open or no ?

@allen_ai Need to try it !

@allen_ai Big win for open earth observation models.

Getting SOTA performance while cutting inference cost by 3x is exactly the kind of efficiency leap that makes real-world deployment actually scalable.

Open + cheaper + same quality is a strong combo.

@allen_ai 🔥🔥

@allen_ai @DFintelligence , utile pour ton projet de localisation de bateaux ? une V2 ?

@allen_ai Thank you @allen_ai for keeping the spirit of openness alive!

@allen_ai nvm found it! if anyone else is wondering it can be found here: https://huggingface.co/datasets/allenai/olmoearth_pretrain_dataset

@allen_ai @Presidentlin

@mmitchell_ai @allen_ai @huggingface Useful reminder that "high impact AI" usually dies on procurement, data rights, and maintenance, not model quality.