Vibeleaderboard
← All Intel
Intel / article

OlmoEarth v1.1: A more efficient family of models

Source
allenai.org
Date

Context

OlmoEarth v1.1 reduces satellite-analysis compute by shortening the sequence the model processes. Earlier versions represented an image patch separately for three groups of spectral bands with different resolutions. Ai2 combined those groups into one representation per patch and time step, producing three times fewer inputs for that imagery. The May 2026 release explains the change.

Simply combining the groups harmed accuracy, so the team also changed . Ai2 reports up to a threefold compute-cost reduction with broadly similar performance across its evaluations, while acknowledging some regressions. The benefit is a lower computational burden for large mapping jobs, with accuracy still dependent on the particular task.

Terms in this piece · Glossary
  • pretraining — The first, biggest phase of building a model: training it on enormous amounts of text so it learns language, facts, and reasoning in general.
  • embedding — A list of numbers representing a piece of text's meaning, so that similar meanings end up numerically close and can be searched.
Read the source allenai.org
Recommended reads
Comments

Checking sign-in…

Loading comments…