In-depth analysis · About 7 min read
Ai2 releases OlmoEarth v1.3 annual embeddings and a browser explorer
OlmoEarth v1.3 makes 10-meter satellite features reusable, with 128- and 64-dimensional options, uneven historical coverage and a restricted-use license.
- Published here
- Source published
- Source checked
Ai2’s October 22 announcement introduces OlmoEarth v1.3 Base, precomputed annual satellite embeddings and an Embeddings Explorer. The release aims to reduce the work needed to turn Earth-observation imagery into useful maps, but its coverage, evaluation results and permitted uses need separate attention.
Listen to the news
Read this article aloud or keep listening to the news.
Preparing the reader…
Device voice
Speech is handled by your browser or device. AISubHub does not send the text to an audio API. Voice availability and playback with a locked screen vary by device.
Resuming may repeat a short phrase.
At a glance
- Ai2 reports that its 128-dimensional embeddings lead 20 of 27 evaluations and the smaller 64-dimensional option leads 18; these are the developer’s results, not AISubHub tests.
- The launch includes global annual coverage for 2023–2025; 2017–2022 cover selected countries. The vectors describe land and adjacent coastal waters, rather than a real-time map of every surface.
- The roughly US$2,720 estimate covers one final pretraining run. Generating a global annual embedding layer is a separate, much larger workload, and the artifact license restricts specified uses.
From repeated model runs to reusable features
OlmoEarth is a remote-sensing model, not a conversational large language model. Ai2’s release combines a Base model with a precomputed archive and a browser interface. Its annual inputs include Sentinel-1 radar and Sentinel-2 and Landsat multispectral imagery. For each 10-meter output pixel, the model produces a vector of 128 numbers; its first 64 dimensions can also be used as a compact representation.
Previously, a typical Studio workflow involved fine-tuning a model for a particular task and then processing satellite imagery. Shared embeddings move much of that feature extraction upstream: a user can train a smaller classifier over existing vectors. That can lower the entry cost for a local experiment, but a vector still needs labels or an analysis method before it becomes a crop classification or change assessment.
How the model compresses a year of observations
The technical report separates three problems: combining observations across time and sensors, controlling inference cost, and keeping the output compact. A vision transformer produces tokens, and an Attention Pooler uses learned location queries to aggregate the observations for each place. A distilled linear projection reduces 768-dimensional location representations to 128 dimensions. Training also supports the first 64 dimensions as a useful nested prefix.
The output resolution is decoupled from the encoder’s patch size. The deployment configuration can encode 20-meter patches while producing embeddings at 10-meter spacing, reducing computation without claiming finer source imagery. The model card lists a 128-million-parameter ViT backbone, a 14.3-million-parameter Perceiver and a 0.1-million-parameter student projection; the backbone count alone is not the size of the entire pipeline.
Read the benchmark claims within their protocols
Ai2’s announcement reports that the 128-dimensional version scores highest in 20 of 27 evaluations against AlphaEarth Foundations and Tessera v2, while the 64-dimensional version leads 18. The evaluations span six continents, and the release also introduces the PLANTEUR tropical crop benchmark. The denominator is the announcement’s evaluation set, not a claim that 27 distinct datasets all produced the same result.
The report examines different classifier choices, task subsets and geographic generalization. Its tests use annual embeddings and the deployed quantization scheme. It also documents cases where adding the Attention Pooler hurts performance, especially some tasks with a single time step or modality. Our reading is that these are promising reusable features, rather than evidence that one configuration wins every mapping task or transfers without local validation.
Coverage and colors are part of the interpretation
At launch, the annual layers for 2023, 2024 and 2025 are global within the released coverage; 2017–2022 are available only for selected countries, with expansion still in progress. The archive focuses on land and adjacent coastal waters and excludes Antarctica. Ai2 describes a full global reference year as roughly 150 TB at 10-meter resolution. The paper also explains how provenance records connect output windows to the contributing scenes and public imagery catalogs.
Ai2 says the Explorer offers year switching, underlying Sentinel-2 imagery, similar-place searches, small classifiers trained from labeled points and GeoTIFF exports without account or GPU setup. Those capabilities were not independently tested here. Its false-color display is a visualization of part of the embedding variation, not a ready-made factual map. A large difference between years signals a changed representation; it does not, by itself, identify mining, regrowth or the cause of a change. Annual aggregation also does not provide real-time alerts.
A small training estimate is not the planetary bill
The announcement estimates roughly US$2,720 for the final v1.3 Base pretraining run on a rented eight-H100 node. The figure uses 539 GPU-hours and an October 2026 Iowa cloud-node rate; it excludes experiments, data preparation, storage and engineering. Ai2 reports training time at 53% of v1.2 and 18% of v1, with estimated emissions at 50% and 22%, respectively. These are remaining fractions of the earlier totals, not the corresponding percentage reductions.
The report puts a separate cost in perspective: generating one global annual embedding layer takes approximately 14,000 H100-hours, an estimate around 26 times the final pretraining run. Actual inference mixes H100 and A100 hardware. AISubHub’s analysis is that precomputation shares a substantial cost across downstream users; it does not erase it. The quoted training estimate is neither a current cloud offer nor a complete budget for reproducing, storing and serving the planetary archive.
Open access still has license and validation boundaries
The model and associated artifacts use the OlmoEarth Artifact License. It grants specified use, modification, distribution and derivative rights subject to conditions, including attribution and preservation of restrictions in downstream distribution. Its prohibited purposes include military and defense applications, intelligence gathering, human surveillance and policing, as well as activities that plan or facilitate extraction of oil, gas or minerals, including mining and deforestation. It should not be described as unrestricted use or substituted with an MIT or Apache license.
The agreement also does not automatically grant every third-party right in dataset contents and supplies artifacts without accuracy or fitness warranties. A team assessing adoption should separate the model license, the rights in its input material, the geographic and annual coverage, and the accuracy needed for its decision. This release supplies building blocks for that assessment; the announcement and its demonstration are not independent proof that a particular exported map is correct.
Why it matters
AISubHub’s analysis: moving the expensive feature extraction into a shared archive can let small teams test mapping questions with fewer local computing requirements. The useful comparison is the full workflow—coverage, labeling, data movement and validation—not the training bill alone.
Limits and availability
AISubHub read the announcement, model card, artifact license and technical report, and inspected the announcement’s figures and short embedded demonstration. We did not download the model or dataset, run the Explorer, reproduce benchmarks, test exports or verify maps against ground truth. Linked external case reports and tutorials were not independently read.
Sources and evidence
Source published: · Source checked:
Further reading
Source checked:
Source checked:
Source checked:
AI-assisted editorial articles checked against linked primary sources. Vendor claims are attributed; inclusion is not endorsement.