Decoding Antiquity is a research initiative launched in the fall of 2025 by ancient historian and papyrologist Anna Dolganov (ÖAI/ÖAW) in partnership with European AI developers Mistral AI and SAIL Reply, David Smith (Northeastern University, Boston, USA), and other international collaborators. The project is dedicated to building industry-grade AI for historical languages and advancing computational research on historical documents, with the explicit aim of aligning AI development with the critical and evidentiary standards of the humanities.

A key aim of the project is to address a fundamental problem in the study of antiquity and wider premodern world, where enormous bodies of written evidence have survived on papyrus, parchment, and other materials, but the vast majority remain unpublished or only partially accessible because their decipherment, reconstruction, and scholarly interpretation are exceptionally labor-intensive and require a high level of specialization. Consequently, archival collections are teeming with texts that have never been read or studied, from inscriptions and papyri to manuscripts and printed books. As an example: of hundreds of thousands of Greek papyri from the Hellenistic, Roman, Byzantine, and early Islamic periods, less than 8% have been deciphered and published. The rest remain a wealth of undiscovered historical knowledge.

The Decoding Antiquity initiative aims to move beyond general-purpose AI to build domain-specific LLMs specialized for historical languages and the specific challenges posed by damaged and fragmentary handwritten documents. To this end, the project is creating Apollo, a family of Large Language Models (LLMs) and Vision-Language Models (VLMs) tailored to different domains of research on historical sources. The vision of Decoding Antiquity is to build an AI-supported research environment that makes historical corpora and previously inaccessible documentary archives legible, navigable, and interpretable, establishing new epistemic parameters for research and opening exciting new possibilities for the study of antiquity and the broader premodern past.

The first model released by the project is Apollo Restore, a foundation LLM for historical Greek trained on ca. 600 million words of premodern and early modern Greek sources and specialized for text restoration. Apollo Restore is designed to propose ranked reconstructions of gaps in damaged, fragmentary, or corrupted Greek texts of all genres, addressing one of the central scholarly tasks in philology, papyrology, epigraphy, and related disciplines. A key feature of Apollo is its ability to propose useful restorations for gaps of unknown length, reflecting realistic conditions of work with fragmentary sources. In its next iteration, Apollo Restore will employ Retrieval Augmented Generation (the RAG for Apollo project), which will enable the model to restore on the basis of parallels and retrieve similar texts across the Greek textual tradition.

The next phase of the project moves from language-based restoration to visual analysis of the documentary record. Work is in progress on multimodal vision-language technology for automated transcription of historical manuscripts, starting with Greek papyri. Documentary papyri present a challenging case for automated text recognition, due to the diversity of cursive writing styles and the fragmentary condition of most of the material. These features pose problems for line-segmentation models used in computer-vision-based approaches. Conventional OCR is also limited in its ability to recover text across physical gaps by drawing on the surrounding textual context. Instead, Decoding Antiquity pursues the path of multimodal foundation models that integrate linguistic knowledge and visual and paleographic evidence and perform whole-page analysis of damaged fragments. A key aim is to design explainable models that provide calibrated uncertainty and explicitly distinguish between attested readings and reconstructed text. 

The agenda of Decoding Antiquity extends beyond Greek to other ancient languages. The Latinitas project in partnership with Mistral AI and SAIL Reply, David Smith, and Neo-Latinist Martin Korenjak of the University of Innsbruck develops a specialized LLM for Latin linked to Greek and RAG systems for cross-lingual retrieval and semantic and thematic search across the Latin and Greek traditions. Future steps will extend the development of domain-specific multimodal LLMs to other large historical languages and explore building advanced AI capabilities for very low-resource languages.

The mission of Decoding Antiquity is inclusive and collaborative. We aim to go beyond the framework of individual projects to join forces in a broader movement to bring the humanities and the study of historical languages and manuscripts into a new digital age. If you are interested in thinking and working together, please contact us!