SSyncDev

OlmoEarth v1.3 Base

SyncDev model directory · Profile updated Nov 17, 2026

Compare with another model

OlmoEarth v1.3 Base is Ai2's open Earth-observation model that turns a year of Sentinel and Landsat imagery into a 128-dimension embedding per 10 m pixel.

Key specifications

Specification score
14.9
License
OlmoEarth Artifact License

The specification score summarizes recorded features and limits. It is not a benchmark result.

Pricing conditions: Weights only; precomputed global embeddings are published separately.

Inputs and outputs

Recorded modality support shows what this model can accept and produce. An undocumented modality is shown as unknown rather than assumed to be unsupported.

Accepts

  • Image input

Produces

  • Embeddings output

Capabilities

VisionOpen weights

What is OlmoEarth v1.3 Base?

Every 10-metre square of land, summarised as 128 numbers per year. That is the output of OlmoEarth v1.3 Base, an Earth-observation foundation model Ai2 announced on 22 October 2026. Feed it a year of satellite time series and it produces an embedding you can cluster, search or train a small classifier on, without touching raw imagery again.

This is not a language model. It belongs with geospatial and image feature extraction tools, and the people who will get value from it map crops, land cover and change.

Inputs and outputs

  • In: one year of Sentinel-1 radar plus Sentinel-2 and Landsat multispectral time series.
  • Out: one 128-dimensional vector per 10 m pixel per year.
  • Smaller option: the first 64 dimensions work on their own when storage matters.

The architecture is a vision transformer followed by an attention pooler that yields one vector per land patch and a linear layer that compresses it to 128 numbers. Ai2 does not publish a parameter count.

You may not need to run it at all

Ai2 published precomputed global annual embeddings at 10 m resolution, in both 64 and 128 dimensions, for 2023, 2024 and 2025. Selected countries go back to 2017, with more planned. The archive is about 150 TB per year, and each 20 km tile lists the exact satellite scenes the model read. A browser explorer needs no sign-in or GPU: compare years, find similar places, train a quick classifier from a handful of labelled points and export a GeoTIFF.

Run the weights yourself only when you need custom dates, modalities or resolution, or a region outside the released coverage.

How it stacks up

Against AlphaEarth Foundations and Tessera v2, across 27 evaluations on six continents, the 128-dimension embeddings rank first on 20, and the 64-dimension version leads on 18. Fine-tuned, it matches v1.2. These are Ai2's own comparisons. Ai2 also introduced a tropical crop-type benchmark, PLANTEUR, with this release.

Cheaper than its predecessor

The final pretraining run would cost about $2,720 of rented H100 time. That is 53% of v1.2's compute and 18% of v1's, with correspondingly lower emissions.

Licence and access

Item Detail
Weights allenai/OlmoEarth-v1_3-Base
Licence OlmoEarth Artifact License (shown as "other" on the hub)
Code, data Pretraining code and data released
Hosted runs OlmoEarth Studio

The licence is Ai2's own, not Apache. Read it before building a commercial product on the embeddings.

A first experiment

Pick one mapping task you already have labels for, crop type or deforestation, say. Download the released embeddings for your area, train a simple linear or tree model on the 64-dimension version, then repeat with all 128. If the small version gets close, you have just cut storage in half. Ai2's other open models are on the Ai2 provider page, and general-purpose vector models sit under embedding models.

How to evaluate OlmoEarth v1.3 Base

Compare the model on the work you need to complete.

Tasks and inputs

OlmoEarth v1.3 Base is listed for Embeddings, Image feature extraction, Embeddings. It accepts image and produces embedding. Use the publisher’s input format and preprocessing for your chosen deployment.

Capacity and cost

Evaluate retrieval over a labelled query set. Keep the encoder, vector dimensions and index settings consistent.

Choosing a deployment

The weights are downloadable under OlmoEarth Artifact License. Check runtime support and size the complete model and working context for your expected concurrency.

Frequently asked questions

One 128-dimensional embedding per 10 m pixel per year, computed from Sentinel-1, Sentinel-2 and Landsat time series. The first 64 dimensions can be used alone.

These records share a directory category. Check the source card, output type and deployment requirements before treating a model as a replacement.

kakao

Kanana-2 Vision Encoder

Kakao's kanana-2-vision-encoder comes in 426M medium and 98M tiny sizes: open native-resolution image encoders for building Korean-strong VLMs.

— context

camenduru

camenduru/dinov3-vitl16-pretrain-lvd1689m

camenduru/dinov3-vitl16-pretrain-lvd1689m is a source-linked Hugging Face repository assigned to Image feature extraction. Its recorded task interface accepts image and produces embeddings.

— context

facebook

facebook/dinov2-base

facebook/dinov2-base is a source-linked Hugging Face repository assigned to Image feature extraction. Its recorded task interface accepts image and produces embeddings.

— context