September 22, 2026
Semantic search comes to papyri: Apollo trained on 600 million words
On September 22, Mistral introduced Apollo, a model for ancient Greek texts. It was trained on roughly 600 million words and tens of thousands of published inscriptions and papyri. Apollo searches by meaning, restores lacunae, and helps decipher manuscripts.
Over 140 years, researchers have published only about 6% of surviving papyri. At the current pace, the full corpus would take more than 2 millennia. Apollo is still being evaluated by international experts, while the next stage will see the Austrian Academy of Sciences add RAG with an open Greek corpus claimed to be 4 times larger than TLG.
The Decoding Antiquity project runs through October 2028, and Apollo and the evaluation results are set to be published openly after review.
Source
