The efficiency problem nobody talks about
When you're running inference across tens of thousands of square kilometers of satellite imagery, compute costs aren't an afterthought—they're the entire budget constraint. Allen AI just shipped OlmoEarth v1.1, a family of remote sensing foundation models that delivers up to 3x lower compute costs while maintaining the same performance benchmarks as the original OlmoEarth release from November 2025.
This isn't a typical model update blog where the team throws more parameters at the problem and calls it progress. It's a methodical rethink of what a token should represent in transformer-based remote sensing models—and why that choice compounds across the entire inference pipeline.
The stakes are real. Organizations using OlmoEarth are tracking mangrove change, classifying forest loss drivers, and generating country-scale crop maps in days. Cutting inference costs by 3x means the difference between running these analyses monthly versus weekly, or scaling from regional pilots to continental deployments.
Why sequence length is the real cost driver
Transformer compute scales quadratically with token sequence length. Everyone knows this in theory, but remote sensing makes it visceral. A single Sentinel-2 input isn't a chat message—it's a tensor with height, width, temporal steps, and 12 spectral channels.
The original OlmoEarth approach split data into spatial patches and created separate tokens for each resolution tier. Concretely: one token per timestep per resolution band. A two-timestep Sentinel-2 input with three resolution tiers (10m, 20m, 60m) yields six tokens per patch. Scale that across an entire image and you're looking at (H/p) × (W/p) × T × 3 tokens total, where p is your patch size.
This design choice wasn't arbitrary. Both Galileo and SatMAE use resolution-specific tokens, and SatMAE's ablation studies showed significantly better results with this separation. The intuition: keeping bands separate makes it easier for the model to learn cross-band relationships, which are critical for interpreting spectral signatures in satellite data.
The 3x question: can you merge tokens without breaking performance?
Here's the trap: if you naively collapse those three resolution tokens into one, you get a 10 percentage point drop on m-eurosat kNN, a standard remote sensing benchmark. Allen AI hypothesizes this happens because merging bands makes it harder to model the cross-resolution relationships that matter for earth observation tasks.
CROMA, another remote sensing model, does use a single token regardless of resolution—so it's not theoretically impossible. But making it work without performance regression required Allen AI to redesign their pretraining regimen. The technical report goes into the specific modifications, but the core insight is that token representation and pretraining objective have to co-evolve.
The result: OlmoEarth v1.1 matches v1 performance on the team's mix of research benchmarks and real-world partner tasks, while cutting MACs (multiply-accumulate operations) by up to 3x. Lower MACs translate directly to faster, cheaper inference.
What this means for practitioners
If you're already running the original OlmoEarth family in production, the migration path is straightforward: try v1.1 on your task. Allen AI is transparent about seeing "some regressions" on specific workloads (check the technical report for details), but if your use case tolerates the tradeoffs, you'll see meaningful speedups during both fine-tuning and inference.
This isn't just about saving cloud credits. Faster inference enables qualitatively different workflows. Planet-scale map refreshes become feasible on tighter cadences. Developers can iterate more quickly during fine-tuning. The OlmoEarth Platform can support more partner organizations with the same infrastructure budget.
Allen AI is shipping model families at three sizes—Base, Tiny, and Nano—so you can pick the compute/performance tradeoff that fits your constraints. The fact that efficiency gains apply across the entire family is the real win here.
The research contribution: isolating methodological changes
Remote sensing models are notoriously hard to study because so many things change at once. New architecture? Different pretraining dataset? Novel augmentation strategy? When performance shifts, attribution is a nightmare.
OlmoEarth v1.1 was trained on the exact same dataset as v1. That means any performance difference isolates the effect of the token representation and pretraining changes. For researchers trying to understand what actually matters when building foundation models for earth observation, this is a controlled experiment at scale.
The field needs more of this. Too many model releases bundle architectural novelty with new datasets and training recipes, making it impossible to disentangle signal from noise. Allen AI's approach—same data, methodological iteration—advances our scientific understanding of remote sensing pretraining in ways that raw benchmark improvements don't.
The bigger picture: efficiency as a product constraint
Foundation model research often optimizes for frontier capabilities: bigger, more accurate, more general. But deployment teams live in a different world where cost per inference and latency percentiles determine what's shippable.
OlmoEarth v1.1 is a model family designed with deployment constraints as first-class concerns. The decision to target 3x efficiency gains—rather than incremental accuracy improvements—reflects a deep understanding of where the real bottlenecks are for remote sensing applications.
This matters beyond earth observation. Every domain with high-resolution spatial or temporal data faces similar tradeoffs. Medical imaging, urban planning, climate modeling—they all deal with large tensors, long sequences, and quadratic attention costs. The lessons from OlmoEarth's token design choices generalize.
Try it yourself
The weights, training code, and technical report are all public. If you're working on remote sensing tasks, the OlmoEarth v1.1 collection on Hugging Face has Base, Tiny, and Nano models ready to download.
Allen AI's broader mission is bringing state-of-the-art AI to organizations working on planetary-scale problems. OlmoEarth v1.1 is a reminder that "state-of-the-art" isn't just about accuracy—it's about making powerful models practical enough to deploy at the scales where they actually matter.