Benchmarking Swedish Lion: Specialized HTR vs. LLMs

Can today's powerful LLMs replace purpose-built HTR models for historical handwriting? The Swedish National Archives put it to the test, benchmarking Swedish Lion against leading language models on 150 manually transcribed pages from 28 archives spanning 1500–1900. The results offer critical, evidence-based guidance for institutions choosing between specialised and generalist AI for archival transcription.