What is EXAONE Path 1.5?
EXAONE Path 1.5 is a whole-slide image analysis framework from LG AI Research. It is designed for digital pathology tasks in which an AI system examines a complete digitized tissue slide rather than a single cropped image. The model can support cancer subtyping, molecular subtyping, and prediction of selected genetic or molecular characteristics from pathology images.
The available release is best understood as a specialized research model. It does not provide a conversational interface, general text generation, or a public chatbot experience. Its output is intended to be a downstream classification result, such as the probability associated with a pathology or molecular prediction.
LG AI Research places EXAONE Path within its broader EXAONE research portfolio. The model is related to the EXAONE Path series, but it addresses a different problem from general-purpose EXAONE language and vision-language models: interpreting large pathology slides for computational pathology research.
How the model processes a whole slide
A whole-slide image can contain a very large amount of tissue information, making it impractical to process the entire image as one ordinary image input. EXAONE Path 1.5 uses a two-stage approach.
- Patch feature extraction: The EXAONE Path 1.0 encoder extracts feature representations from smaller patches sampled from the whole-slide image.
- Slide-level aggregation: A Vision Transformer-based module combines the patch features into a representation of the complete slide.
- Downstream classification: A linear classifier uses the slide-level representation to produce a task-specific prediction.
In plain terms, the system first turns many local tissue regions into numerical descriptions and then learns how those regions should be considered together. This is important because a mutation-related signal or tumor subtype may depend on patterns distributed across multiple regions of a slide, not just on one isolated patch.
During pretraining, LG AI Research used multimodal learning to align slide images with mRNA gene-expression profiles. That research approach connects visual pathology patterns with molecular information. However, the released EXAONE Path 1.5 downstream model should not be described as a general multimodal assistant: its documented practical use is whole-slide image inference for pathology tasks.
Documented use cases
The primary use case is retrospective or experimental analysis of digitized pathology slides. The supplied model information identifies the following applications:
- Mutation prediction from whole-slide pathology images
- Molecular and tumor subtyping
- Research into image-based precision oncology
- Evaluation of pathology representations for downstream machine-learning tasks
The released downstream checkpoint is specifically associated with EGFR mutation prediction in lung adenocarcinoma, often abbreviated as LUAD. That specialization matters: a result for this task should not automatically be interpreted as evidence that the model works equally well for other cancers, mutations, tissue types, or clinical settings.
Reported performance and hardware requirements
The EXAONE Path 1.5 model card reports an area under the receiver operating characteristic curve (AUC) of 0.81 for LUAD EGFR mutation prediction and an average AUC of 0.76 across the listed evaluation tasks. AUC is a ranking metric that indicates how well a model separates two classes across possible decision thresholds; it is not the same as clinical accuracy, and it does not by itself establish clinical usefulness.
These figures are reported model-card results for the stated evaluation settings. They should not be treated as a guarantee of performance on slides from a different hospital, scanner, staining process, patient population, or data-preparation pipeline. Independent validation is especially important in pathology because changes in slide preparation and acquisition can affect model inputs.
The implementation requires an NVIDIA GPU, with approximately 40 GB of GPU memory recommended. This makes the model more suitable for a research workstation, laboratory server, or managed on-premise environment than for casual experimentation on an ordinary laptop. The supplied information does not specify a context window, maximum token output, image-dimension limit, or universal slide-size limit. Those values should therefore be treated as unknown rather than inferred from the model architecture.
Inputs, outputs, and capabilities
| Capability | Verified position |
|---|---|
| Primary input | Whole-slide pathology images, processed through image patches |
| Primary output | Downstream classification probabilities or related task predictions |
| Text generation | Not supported as a documented model function |
| Image generation | Not supported |
| Audio or video | Not supported |
| Tool or function calling | No documented support |
| Web search | No documented support |
| Fine-tuning | Fine-tuning scripts are provided for research use |
| Hosted API | No official hosted inference API was identified |
Although the wider EXAONE family includes multimodal research, EXAONE Path 1.5 should be evaluated as a specialized pathology vision system. It is not a text-and-image assistant that accepts questions about a slide and returns a written medical explanation. Users may need to build their own preprocessing, inference, evaluation, and reporting workflow around the model.
Access, licensing, and pricing
EXAONE Path 1.5 is distributed through its Hugging Face model repository, subject to the repository's access conditions. Access requires acceptance of the applicable conditions and sharing contact information. The model is licensed under the EXAONEPath AI Model License Agreement 1.0 - NC.
No official hosted API or token-based pricing was identified for this model. There is therefore no verified per-request, per-image, monthly subscription, or commercial inference price to report. The practical cost of using it may instead include GPU infrastructure, storage for large slide files, data preparation, engineering work, and any obligations imposed by the model license or a research institution.
The license and access conditions should be reviewed before commercial, clinical, or redistribution use. Availability of model weights does not by itself establish permission to use the system for diagnosis or to deploy it in a regulated medical workflow.
Main strengths and limitations
Strengths
- Whole-slide focus: The architecture is designed to aggregate information across an entire pathology slide rather than treating each image patch as an independent final decision.
- Pathology specialization: Its training and intended tasks are closely aligned with computational pathology and molecular oncology research.
- Molecularly informed pretraining: The use of slide images and mRNA gene-expression profiles during pretraining provides a research connection between visual and molecular pathology information.
- Adaptability for research: Fine-tuning scripts and an open research release make it possible for qualified teams to investigate additional downstream tasks, subject to the license and available data.
Limitations
- Narrow task scope: The released downstream checkpoint targets EGFR mutation prediction in lung adenocarcinoma rather than a broad range of clinical questions.
- Research status: It is not presented as a standalone clinical diagnostic system, and the supplied information does not establish regulatory approval or clinical validation.
- Deployment burden: The recommended NVIDIA GPU memory of about 40 GB and the need to process large slides create meaningful infrastructure requirements.
- No turnkey service: There is no identified official hosted inference API, consumer application, or token-based pricing plan.
- Unknown generalization: Performance may differ across institutions, scanners, staining protocols, and patient populations. Additional evaluation and fine-tuning may be necessary.
When to choose EXAONE Path 1.5
Choose EXAONE Path 1.5 when the project involves whole-slide digital pathology and the team can operate a GPU-based research workflow. It is a reasonable candidate for researchers investigating image-based molecular prediction, tumor classification, or precision-oncology methods, particularly when access to the model's pathology-specific representations is more valuable than a simple hosted endpoint.
It may also be appropriate when the team needs to experiment with fine-tuning on institution-specific pathology data. In that situation, the model should be treated as a starting point for research rather than as a ready-made diagnostic product. A careful evaluation should include representative local data, a clear train-validation-test design, calibration analysis, and review by pathology experts.
Another type of option may be more appropriate when the requirement is a general medical-imaging platform, a managed inference service, a conversational report generator, or a model with established clinical deployment support. A general-purpose vision-language model may be easier to interact with, but it is not automatically a substitute for a pathology-specific whole-slide model. Conversely, a commercial digital pathology platform may offer more operational support while providing less flexibility for model research.
Overall assessment
EXAONE Path 1.5 occupies a specialized position: it is a research-oriented whole-slide pathology framework for converting patch-level image information into slide-level molecular or cancer-related predictions. Its strongest distinction is not conversational breadth or API convenience, but its focus on pathology representation learning and downstream analysis.
The model is most compelling for research teams that have suitable whole-slide data, GPU resources, and the expertise to validate results. It is a poor fit for users seeking a general chatbot, an inexpensive hosted API, or an immediately deployable clinical diagnostic tool. The reported LUAD EGFR results provide a useful reference point, but safe use requires independent validation and strict attention to the model's task, license, and clinical limitations.