OlmoEarth v1.1 slashes satellite AI compute by 3x by killing redundant tokens

The team behind OlmoEarth just made a move that’s more about smart engineering than raw power, and it could quietly reshape who gets to use planetary-scale AI. Their new v1.1 model family cuts compute costs by up to three times compared to the original November 2025 release, without tanking performance. That’s a big deal when you’re processing satellite imagery across entire continents.

The efficiency gain comes from a fundamental redesign of how the model sees data. Transformer models like OlmoEarth chew on tokens, and the cost scales quadratically with the sequence length. The old approach created separate tokens for each spatial resolution band in Sentinel-2 imagery—10m, 20m, and 60m—which ballooned the token count. Simply smashing them into one token was a non-starter; the team saw a brutal 10 percentage point drop on the m-eurosat kNN benchmark. So they didn’t just collapse tokens. They changed the pre-training regimen to teach the model to understand cross-band relationships without the crutch of individual resolution tokens.

The result is a leaner model that sidesteps a trade-off that researchers at Galileo and SatMAE previously accepted as necessary. As the team puts it, “A more efficient model means we can support more partners on the OlmoEarth Platform.” This isn’t an abstract research win. The original OlmoEarth was already being used for tracking mangrove change and generating country-scale crop maps in days. Cutting the largest cost center—inference compute—directly translates to more frequent map refreshes for organizations that can’t afford to burn cash on GPUs.

I’m curious about those “some regressions” the team openly cops to in their technical report. A 3x speedup is phenomenal, but for precision agriculture or disaster response, even a small accuracy slip on a niche task matters. The real test will be whether partners adopt v1.1 for their specific workflows or stick with the proven, pricier original. For researchers though, this is a clean experiment: same dataset, different method. That kind of controlled comparison is rare and useful for a field where it’s often impossible to tell if a model is better because of the architecture or just because it saw more data.