OlmoEarth-v1-Large

What is OlmoEarth-v1-Large?

OlmoEarth-v1-Large is an open Earth-observation foundation model developed by the Allen Institute for AI. It uses a vision transformer, a neural-network architecture designed to analyze visual data, to learn representations from satellite imagery and image time series. Those representations can then be reused by other models or task-specific systems.

The model is not a conversational AI assistant and does not generate ordinary written answers. Its primary output is an embedding: a numerical representation of an image or sequence of images that captures information useful for comparing observations, training classifiers, segmenting regions, or building geospatial prediction systems.

Ai2 identifies OlmoEarth-v1-Large as part of the original OlmoEarth v1 family, alongside Nano, Tiny, and Base variants. It is the largest checkpoint in that family. Later OlmoEarth v1.1 and v1.2 releases are separate model families and should not be treated as aliases or updates to this specific checkpoint.

Architecture and training approach

The checkpoint has approximately 308 million encoder parameters and 53 million decoder parameters. The encoder is the component most directly associated with extracting useful features from satellite observations, while the encoder-decoder training design helps the model learn from partially masked inputs and related prediction targets.

OlmoEarth-v1-Large was trained using masked image modeling and contrastive objectives. In masked image modeling, portions of an input are hidden and the model learns to infer information about them. Contrastive learning encourages representations of related observations to be appropriately similar or different. Together, these objectives are intended to produce reusable features rather than a narrow prediction for only one labeled task.

The training setup combines three important Earth-observation sources: Sentinel-1, Sentinel-2, and Landsat. Sentinel-1 contributes radar observations, while Sentinel-2 and Landsat provide optical and multispectral information. The model also uses observations collected across time and derived geographic products as prediction targets. This multimodal and multitemporal design is useful when a task depends on both what is visible in an image and how a location changes over time.

Supported inputs and outputs

The primary supported inputs are Sentinel-1 imagery, Sentinel-2 imagery, Landsat imagery, and image time series based on those observations. In practical terms, users need to prepare satellite data in a way that matches the model’s expected sensor, spatial, spectral, and temporal conventions. Performance can depend substantially on preprocessing and sampling choices.

OlmoEarth-v1-Large produces learned image representations and embeddings for downstream use. It does not natively produce text, generated images, audio, video, speech, or music. It also has no reported conversational context window or maximum text-output limit because it is not a language model and does not expose a standard token-based generation interface.

What can it be used for?

The model is designed to serve as a reusable representation-learning component in Earth-observation systems. A user can extract embeddings from satellite observations and provide them to a downstream classifier, segmentation model, search system, or geospatial analytics pipeline. The model can also be fine-tuned for a particular task when labeled data and suitable training infrastructure are available.

• Land-cover and land-use classification: distinguish categories of terrain or human activity using satellite-derived features.
• Segmentation and mapping: identify or delineate geographic regions and objects.
• Agricultural analysis: support crop-related classification and monitoring workflows.
• Ecosystem and environmental monitoring: analyze geographic conditions and changes over time.
• Change analysis: compare observations from different dates to help identify meaningful changes.
• Embedding search: represent satellite images numerically so similar observations can be retrieved or grouped.
• Geospatial fine-tuning: adapt the pretrained representation to a domain-specific prediction task.

These are appropriate use cases for an embedding-focused remote-sensing model. They should not be confused with a claim that the checkpoint independently completes a full production workflow. A downstream task head, fine-tuning process, preprocessing pipeline, or separate geospatial application may still be required.

Main strengths and trade-offs

The most important strength of OlmoEarth-v1-Large is its specialization. Rather than treating satellite data as generic images, it is trained around Earth-observation inputs from multiple sensors and across time. Radar, optical, and historical observations can expose different aspects of a location, so this design may be more useful for geospatial representation learning than a general computer-vision model that was not trained on comparable data.

Its scale is another practical advantage within the original OlmoEarth v1 family. The approximately 308-million-parameter encoder offers a larger representation-learning capacity than the smaller Nano, Tiny, and Base variants. The trade-off is that a larger checkpoint generally requires more memory and computational resources to run or fine-tune. The supplied research does not provide a benchmark table quantifying the speed or accuracy difference between these variants, so any claim of a specific performance advantage would be unverified.

Editorially, the model is best viewed as a high-capability option for demanding Earth-observation embedding and fine-tuning workloads, with a corresponding cost in hardware and operational complexity. The supplied capability scores rate its intended remote-sensing usefulness highly, but those scores are editorial assessments rather than provider-published benchmark results.

When to choose OlmoEarth-v1-Large

Choose OlmoEarth-v1-Large when the central problem involves satellite imagery, especially when the workflow can benefit from multiple sensors, repeated observations, and a pretrained representation. It is a strong candidate for teams that need a large remote-sensing encoder for classification, segmentation, environmental mapping, or embedding extraction and are prepared to manage model execution and downstream task development.

Its open-model distribution is also relevant when an organization needs access to a checkpoint and training resources rather than a conventional hosted language-model API. Ai2 provides the model through its official Hugging Face organization, along with training code, evaluation resources, tutorials, and access through the OlmoEarth Platform.

A smaller OlmoEarth v1 checkpoint may be more appropriate when hardware capacity, inference speed, or operating cost matters more than the maximum representation capacity available within the original family. The research confirms that Nano, Tiny, and Base variants exist, but it does not provide enough comparative measurements to select one based on a specific accuracy or latency threshold.

A general-purpose vision model may be a better fit for ordinary photographs or image-generation tasks, while a language model is more appropriate for conversation, code generation, document processing, or text output. OlmoEarth-v1-Large is not a substitute for those model types. Even within Earth observation, another model may be preferable if its data coverage, preprocessing requirements, license, task-specific benchmarks, or deployment format better match the project.

Pricing, access, and license

There is no official hosted token pricing for OlmoEarth-v1-Large in the supplied information. The checkpoint is distributed as an open model, so the cost structure differs from a typical API model with per-input and per-output token rates. Users may incur infrastructure costs for storage, hardware, inference, fine-tuning, or platform access. Terms for the OlmoEarth Platform may apply separately.

The model is distributed under the OlmoEarth Artifact License. The supplied research specifically notes restrictions related to military, defense-related, and extractive-industry applications. Organizations should review the license before deployment rather than assuming that open checkpoint availability permits every commercial or operational use.

Limitations and deployment considerations

OlmoEarth-v1-Large does not provide a general-purpose hosted chat API, standard token context window, or normal language-model output limit. It cannot directly answer questions in natural language, call external tools, return generated images, or process speech as a native conversational system.

Its practical results depend on more than parameter count. Sensor preprocessing, spatial resolution, temporal sampling, geographic coverage, downstream task design, hardware, and available labeled data can all affect the outcome. An embedding is a useful intermediate representation, not a complete application or guaranteed prediction.

Users should also account for the model’s computational demands. The large checkpoint may be a poor fit for constrained edge hardware or high-throughput workloads where a smaller model offers an acceptable quality-to-cost balance. The supplied sources do not specify a universal memory requirement, inference speed, maximum image size, or supported time-series length, so those values should be verified in the official model documentation before implementation.

Bottom line

OlmoEarth-v1-Large is a specialized, large-scale foundation model for satellite-image representation learning. Its value comes from learning across Sentinel-1, Sentinel-2, and Landsat observations, including multitemporal data, and making those learned features available for geospatial tasks. It is most suitable for teams building Earth-observation classifiers, segmentation systems, mapping tools, environmental analyses, or embedding pipelines. It is not suitable when the requirement is text generation, conversational interaction, media generation, or a conventional metered LLM API.