OlmoEarth v1.2 switches to RoPE to reduce embedding artifacts

OlmoEarth is Ai2's family of open models for satellite and other Earth-observation data. Many users take its embeddings, the lists of numbers it produces to summarize each patch of an image, and feed them into their own maps or classifiers. A model like this has to know where each patch sits in space and time. Earlier versions did that by adding a position code to each patch's numbers, and the technical report says that code could leak through into the final embeddings, producing striping aligned with the image grid when large areas were processed tile by tile.

Version 1.2, released in June 2026, uses rotary positional embeddings instead. Position is no longer added to the patch data. It is applied inside the attention step, where the model compares patches: the vectors used in that comparison are rotated by amounts that depend on each patch's location and time step. Codes for month and data source are still added as before. The report shows the stripes disappearing in visualizations but gives no numerical measure of the artifacts.

Ai2 calls the performance gain small, and the report supports that wording. In its direct comparison, rotary embeddings beat the old method on some benchmarks, such as PASTIS and BigEarthNet, and lost on others, including MADOS and EuroSAT. Version 1.2 also adds a Small model between Tiny and Base. It follows the v1.1 update, which cut compute by merging image tokens, and the report describes the changes from both releases together.