Austrian Academy of Sciences releases Apollo, first advanced LLM built for Ancient Greek

The Austrian Academy of Sciences will release Apollo on Wednesday, the first advanced large language model developed specifically for Ancient Greek. Built in partnership with the French AI lab Mistral and the technology-services company Sail Reply, the model is available to researchers at no cost through a chatbot interface. The stated goal: dramatically cut the time needed to identify relevant passages among hundreds of thousands of fragmentary papyri held in academic libraries worldwide, and to offer statistical completions for the missing sections.

Apollo was trained on a corpus of roughly 600 million words of historical Greek drawn from manuscripts, papyri and inscriptions. Its core technical advantage, according to Anna Dolganov, a historian and papyrologist at the Austrian Academy, is the ability to recognise linguistic context without explicit instruction: when it encounters Homeric text it completes in Homeric Greek; when it detects a Doric inscription it switches to Doric. That capability replaces a complex manual stage in which scholars had to determine word boundaries — Ancient Greek writes without spaces — date the document, and weigh socio-political context before attempting any completion.

Reconstructing a torn papyrus has until now required rare expertise. Stephen Colvin, a linguist and historian of classics at University College London, notes that very few people worldwide command the necessary knowledge. Dimitris Vlitas, a partner at Sail Reply, told WIRED that such a capability "was unthinkable a year ago." Armand D'Angour of Oxford, which holds the world's largest papyrus collection, estimates that even offering three candidate words for a single gap would significantly accelerate the workflow and free researchers to focus on historical interpretation rather than technical decipherment.

Despite the enthusiasm, scholars are quick to bound expectations. Colvin stresses that most unreconstructed papyri are everyday documents — personal letters, marriage contracts, bureaucratic paperwork — not lost plays of Sophocles. D'Angour adds that each completion contributes a "tiny fragment of knowledge" about antiquity; together they may confirm existing hypotheses or reveal new details of daily life. The change will not make headlines, but will accumulate in the information layer.

The inherent risk of a probabilistic model is contaminating the historical record with errors that look plausible. To mitigate this, Apollo is programmed to present a set of candidate completions rather than a single answer, leaving the final judgment to the researcher. If the model proves itself, Vlitas says, the same methodology can be transferred to other ancient languages — Latin, Egyptian — or to any academic field wrestling with a large, unmapped corpus. Meanwhile, AI has already logged parallel breakthroughs: OpenAI reported solving a 200-year-old mathematics problem, and Google DeepMind published an extensive map of genetic mutations in molecular biology.