The latest iteration of the OlmoEarth model, version 1.1, has been unveiled, promising enhanced efficiency for Earth observation tasks. Released by Allen Institute for AI (AI2) on May 19, 2026, this model builds upon its predecessor, OlmoEarth v1, which debuted in November 2025.
Since its initial launch, OlmoEarth has been utilized in diverse applications, from monitoring mangrove ecosystems to mapping crop types at a national scale. Its capacity to process satellite imagery over vast areas has been pivotal in supporting organizations dedicated to environmental protection.
Efficiency and Cost Reduction
OlmoEarth v1.1 introduces a family of models that can reduce compute costs by up to 3x compared to the original version, all while maintaining performance across various research benchmarks. The efficiency gains are crucial, as the computational demands of running these models can be significant, particularly during data export, preprocessing, inference, and post-processing phases.
In transformer-based models like OlmoEarth, two primary factors influence efficiency: model size and token sequence length. The latter is particularly impactful, as compute costs increase quadratically with longer token sequences. By optimizing these parameters, OlmoEarth v1.1 enables users to deploy the technology more affordably and swiftly.
Token Design and Model Performance
One of the pivotal design choices in OlmoEarth v1.1 involves how tokens are defined for processing remote sensing data. For instance, when working with Sentinel-2 imagery, the model converts the data into a sequence of tokens based on spatial patches. This method allows for the creation of multiple tokens per patch, depending on the resolution and temporal dimensions.
While traditional methods have employed separate tokens for different resolutions, OlmoEarth v1.1 explores the potential of merging these tokens. This approach aims to reduce the overall token count, thereby enhancing efficiency without significantly compromising model performance. However, it is noted that some performance regressions have occurred, which are documented in the accompanying technical report.
Implications for Researchers and Developers
For developers currently using the original OlmoEarth model, transitioning to v1.1 could yield substantial benefits, including faster fine-tuning and inference times. Researchers will find that the consistent training dataset between versions allows for clearer insights into the impacts of methodological changes on model performance.
In summary, OlmoEarth v1.1 represents a significant step forward in the realm of Earth observation, balancing the need for efficiency with the imperative of maintaining high performance across a variety of tasks. The model’s advancements are now available for exploration, with weights and training code accessible for further development.
This article was produced by NeonPulse.today using human and AI-assisted editorial processes, based on publicly available information. Content may be edited for clarity and style.
