Our Take
Three-fold efficiency gains on production workloads matter more than benchmark noise; the real win is affordability for planet-scale teams running this daily.
Why it matters
Satellite teams pay for inference at scale (data export, preprocessing, inference, post-processing). A 3x cost cut means more frequent continental and national-level map refreshes on the same budget, opening deployment to smaller organizations working on mangrove tracking and crop classification.
Do this week
Remote sensing teams: benchmark OlmoEarth v1.1 against your v1 fine-tuned weights this week so you can quantify speedup and cost savings before production swap.
Allen Institute ships 3x faster satellite models via token redesign
Allen Institute released OlmoEarth v1.1, a family of transformer-based models for remote sensing that cut compute costs by up to 3x while maintaining performance parity with OlmoEarth v1 across internal benchmarks and partner tasks (company-reported). The improvement comes from a single architectural change: merging multi-resolution tokens into single tokens per patch.
OlmoEarth processes satellite imagery (Sentinel-2 and other modalities) by converting it into token sequences. The original v1 design created one token per timestep per resolution (10m, 20m, 60m bands yielded 3 tokens per patch per timestep). This approach, used by competitors like SatMAE, was believed necessary to preserve cross-band relationships. OlmoEarth v1.1 collapses all bands into one token per patch per timestep, reducing token count by a factor of three.
Naive merging caused a 10 percentage-point drop on m-eurosat kNN, a standard remote sensing benchmark. Allen Institute modified the pre-training regimen to recover performance; the resulting model matches v1 on their test suite while requiring one-third the compute. The changes are detailed in their technical report.
The efficiency gain translates directly to inference cost. Multiply-accumulate operations (MACs) scale quadratically with sequence length; even small reductions compound across preprocessing, fine-tuning, and inference. For teams running satellite predictions across tens to hundreds of thousands of square kilometers, compute dominates the cost structure.
Allen Institute has released model weights for Tiny, Base, and Nano sizes, along with training code and pre-training implementation on Hugging Face. Partners have already deployed OlmoEarth v1 for mangrove change tracking, forest-loss classification, and country-scale crop-type mapping.
Cost efficiency unlocks frequency and scale for real-world deployments
Satellite teams do not run inference once; they refresh maps continuously as new imagery arrives. A 3x cost reduction means the same budget supports quarterly or monthly updates instead of annual ones, or the same refresh cadence at one-third the expense.
For organizations like conservation groups, government agencies, and agricultural firms, this cost floor matters more than marginal accuracy gains. OlmoEarth v1.1 does not claim new capabilities; it performs at v1's level. The value is affordability at production scale. Teams running the platform can support more partners, and practitioners deploying on their own infrastructure see faster inference and lower hardware bills.
Allen Institute also separated the dataset (identical to v1) from the methodology, creating a controlled comparison for researchers. Any performance difference between v1 and v1.1 isolates the effect of token design and pre-training changes, advancing understanding of what actually matters when training remote sensing transformers.
Audit your OlmoEarth deployment for speedup
If you are running OlmoEarth v1 in production, test v1.1 on your fine-tuned weights before migrating. The technical report flags some regressions on specific tasks; whether v1.1 works depends on your use case. If it does, you inherit one-third compute cost and measurably faster fine-tuning and inference without retraining from scratch.
For new projects, start with v1.1 and the Nano or Tiny models if your hardware budget is constrained. The model family approach means you can tune to your compute envelope without sacrificing inference speed.