OlmoEarth v1.1: A more efficient family of models

The Allen Institute for AI (AI2) has released OlmoEarth v1.1, an updated family of foundation models designed for geospatial analysis. This new version focuses on computational efficiency, achieving up to a threefold reduction in compute costs compared to its predecessor, OlmoEarth v1, while maintaining comparable performance. The update is significant for organizations that process vast amounts of satellite imagery for applications like environmental monitoring and agriculture, as lower compute requirements make planet-scale analysis more accessible and affordable.

Reducing Tokens for Greater Efficiency

The efficiency gains in OlmoEarth v1.1 stem from a fundamental change in how the transformer-based model processes satellite data. AI2 engineers re-architected the model's tokenization strategy for multi-resolution imagery, such as that from Sentinel-2 satellites. Because compute costs in transformers scale quadratically with the input token sequence length, reducing the number of tokens provides substantial savings.

• The original v1 model created a unique token for each of Sentinel-2's three resolutions (10m, 20m, 60m), resulting in three tokens per data patch per timestep.
• The new v1.1 model collapses these resolutions into a single, unified token. This simple change reduces the total token count by a factor of three.
• To avoid the performance degradation that typically accompanies such a change, AI2 modified its pre-training regimen to help the model learn cross-band relationships within a single token.

Implications for Geospatial AI

For developers and partners using the OlmoEarth Platform, the v1.1 models directly translate to faster, cheaper fine-tuning and inference. This makes it more feasible to conduct frequent, large-scale map refreshes for tasks like tracking crop types or forest loss. For the research community, this release offers a valuable case study. Since OlmoEarth v1.1 was trained on the same dataset as v1, it isolates the impact of tokenization and pre-training methods, providing clearer insights into building efficient models for remote sensing. The weights for the Base, Tiny, and Nano models are publicly available.

Strategic Takeaway: AI2's work on OlmoEarth v1.1 demonstrates that significant efficiency gains in specialized foundation models can come from intelligent data representation and pre-processing, not just from shrinking model size. Optimizing the tokenization strategy is a critical and sometimes overlooked lever for reducing the high computational costs associated with scaling transformer models for domain-specific data like satellite imagery.