Atlantes

What is Atlantes?

Atlantes is a suite of GPS transformer models created by the Allen Institute for AI. Unlike a conventional language model, it does not primarily read and write text. Instead, it analyzes sequences of location and movement data from vessels. The system is intended to turn raw Automatic Identification System (AIS) messages into useful maritime intelligence.

AIS is a tracking system used by ships to broadcast information such as position, movement and other navigation-related data. Atlantes applies machine-learning techniques to these sequences to perform two main tasks: entity classification and vessel-activity classification. In practical terms, it can help determine which vessel or vessel type is represented by a sequence and whether the vessel is transiting, anchored, fishing, moored or engaged in another activity.

The release is open source and includes model weights, training code, inference code and production service code. This makes Atlantes primarily a deployable research and operational system rather than a hosted chatbot or general-purpose AI API.

Purpose and position in the Allen Institute for AI catalog

Atlantes occupies a specialized position in the Allen Institute for AI's catalog. Its target data is continuous GPS or AIS telemetry, and its outputs are classifications that support maritime analysis. It is therefore better understood as a domain-specific sequence model than as a language model.

The provider describes Atlantes as a system for real-time, global-scale maritime intelligence. Its intended applications include monitoring vessel activity, identifying possible illegal fishing, supporting marine conservation and researching GPS trajectories. The model was trained using more than 15 million expert-labeled AIS messages, according to the project materials.

The architecture uses continuous point embeddings, convolutional layers and transformer encoders. These components allow the system to represent individual GPS points and analyze patterns across a longer movement sequence. A transformer here is being used to model relationships within a sequence of locations and observations, not to generate a natural-language response.

How Atlantes processes maritime data

Atlantes accepts raw AIS or GPS sequences rather than ordinary text prompts. The implementation supports a configurable context of up to 2,048 GPS messages, spanning approximately 30 days in the described use case. This limit refers to the number of sequence observations or messages, not to a token window measured in words or subword units.

The release contains separate entity and activity classifiers. An entity classifier can help identify the vessel or entity represented by the trajectory, while an activity classifier predicts what the vessel is doing. Reported activity categories include:

• Transiting
• Anchored
• Fishing
• Moored
• Other activities

The published implementation is designed to process activity classifications in real time on modest hardware. That emphasis is important: Atlantes is intended for ongoing monitoring pipelines in which new position messages arrive continuously, rather than for occasional offline analysis alone.

Capabilities and limitations

Atlantes' principal capability is sequence classification over maritime movement data. It can support systems that flag patterns for further investigation, organize vessel trajectories, or classify activity at scale. Its released training and inference code also gives technical users a basis for adapting the system to their own data or classification needs.

The model does not provide the capabilities associated with a general-purpose language model. It is not designed to write reports, answer questions in natural language, summarize documents, generate software, interpret images or process audio. The research describes no multimodal image, audio or video input, and no direct text-generation output.

There is also no documented hosted API with standard per-token input or output pricing. Users should expect to deploy and operate the open-source implementation themselves. That can reduce dependence on a commercial inference service, but it transfers responsibility for hardware, data pipelines, monitoring, scaling and maintenance to the deploying organization.

The published context limit is 2,048 GPS messages. No separate maximum output-token limit is documented because Atlantes produces classification results rather than a long-form generated response. The supplied materials also do not document a general tool-calling interface, structured natural-language output mode, streaming API, batch API or caching service.

Pricing and access

Atlantes has no official hosted API pricing in the supplied documentation. It is released as an open-source, self-hosted model suite rather than as a conventional paid text-model endpoint. The model weights and implementation are available through the project's official repository and release materials.

“Open source” does not mean that deployment is cost-free. An organization may still incur costs for compute, storage, AIS data access, engineering, operations and integration with existing maritime systems. The practical cost advantage is that there is no documented provider charge per input or output token and no requirement to use a hosted conversational API.

Main strengths

• Specialized input representation: Atlantes is built for GPS and AIS sequences rather than forcing maritime movement data into a text-oriented format.
• Real-time orientation: The implementation is designed for activity classification as new vessel messages arrive and is reported to run on modest hardware.
• Operationally relevant labels: The activity categories include fishing, anchoring, transiting and mooring, which map directly to maritime-monitoring workflows.
• Open implementation: Model weights, training code, inference code and production service code are included in the release.
• Adaptability: Fine-tuning is supported through the released training code, allowing technical teams to investigate customization for their own labeled trajectory data.
• Large labeled training set: The project reports training on more than 15 million expert-labeled AIS messages.

Reasoning, coding and supported modalities

Atlantes should not be evaluated using the same capability categories commonly applied to chat models. It has no conversational reasoning mode and does not generate code. Its useful “reasoning” is specialized pattern recognition over sequences of vessel positions and related AIS observations.

Its input is structured GPS or AIS sequence data. Its output is classification information produced by the model pipeline. It does not directly generate text, images, audio or video. There is no evidence in the supplied research of image, audio or video input support, and no documented function-calling or tool-use layer.

Fine-tuning is supported in the practical sense that the released training code can be used to adapt the model. This should not be confused with a managed fine-tuning product: the available research does not describe a hosted training service, automated job API or provider-operated deployment.

Speed and cost trade-offs

Atlantes is optimized for a different trade-off from large general-purpose models. Its specialization and relatively focused classification tasks can make it more suitable for low-latency maritime monitoring than a much larger model that would first need to interpret movement data indirectly. The project specifically describes real-time processing on modest hardware.

Its cost profile is also different from a hosted API. Self-hosting avoids documented per-token charges, but it requires technical infrastructure and operational work. For a research group or maritime organization that already manages AIS data and needs repeatable local inference, that trade-off may be attractive. For a user who only needs occasional natural-language answers, self-hosting a GPS transformer would be unnecessary.

These are practical evaluations based on the model's documented design, not provider-published benchmark scores. The supplied research does not provide a comparative benchmark showing that Atlantes is faster or more accurate than a named alternative across a common test set.

Best use cases

Atlantes is a strong fit when the central problem involves vessel trajectories and activity classification. Suitable applications include:

• Real-time classification of vessel activity from AIS feeds
• Illegal-fishing monitoring and investigation support
• Marine conservation and protected-area surveillance
• Research on GPS trajectories and maritime behavior
• Vessel-activity indexing for large historical datasets
• Local or private maritime-intelligence deployments where raw tracking data should remain under the operator's control

Its output could form one component of a broader system that stores trajectories, applies business rules, sends alerts or presents evidence to human analysts. The model itself should not be treated as a complete enforcement or decision-making system; classification results may need validation, additional context and expert review.

When to choose Atlantes

Choose Atlantes when you need an open-source model specifically designed to classify vessel entities or activities from AIS/GPS sequences, especially when real-time processing, local deployment and the ability to fine-tune the released code matter. It is particularly relevant for maritime researchers, conservation organizations and engineering teams building custom vessel-monitoring pipelines.

Another type of model is more appropriate when the primary requirement is text generation, question answering, report writing, code assistance, document analysis or multimodal interpretation. A hosted language-model API may also be preferable when the organization does not want to operate its own inference infrastructure. Similarly, if the task requires images, audio, video or natural-language interaction, Atlantes is not the appropriate component based on the documented capabilities.

Bottom line

Atlantes is a focused, open-source GPS transformer suite for maritime intelligence, not a general-purpose chatbot. Its defining value is the ability to classify vessel movement and activity from AIS sequences, with a context of up to 2,048 GPS messages and an implementation aimed at real-time use on modest hardware. The main decision is whether that specialized capability justifies self-hosted deployment. For maritime monitoring and trajectory research, it offers a relevant foundation; for language, coding or multimodal applications, a different model category is needed.