This release includes a number of major changes that transform Leo from a transcription tool into a comprehensive platform for digital scholarship
Major changes
Introduced multi-model support, allowing users to transcribe using several models in addition to Leo’s proprietary ATR-1, including GPT-5.2, Gemini 3 Flash, Claude Opus 4.5
Introduced transformation tools, a feature for advanced analysis that reworks transcription output into annotations, with options to correct, modernize, translate, summarize, extract named entities from transcriptions, and more
Introduced tabs to organize multiple transformations and annotations, including the ability to compare multiple transcriptions and interpolate between them
Minor changes
Updated the search feature to handle tabular content, including the option to only search primary transcriptions
Updated sharing links to specify which tabs to share
Made the Intercom widget collapsible to avoid overlap during image list navigation
Made the tasks widget collapsible and moved it to the toolbar with a count icon
Made it possible to shift-select items in the image list
Stopped transcripts collapsing multiple spaces in a row upon saving
Introduced inline renaming for the image list viewer
Fascinating! How do you evaluate ATR-1 in comparison to the others, though? In my experience, I generally prefer ATR-1 to others, even more advanced than the ones offered, such as Gemini 3.1 Pro (which seems to be the one with best computer vision, although I haven’t tested GPT-5.4 yet).
Hey Thiago! We have yet to run a formal validation study, but in my experience of testing different kinds of manuscripts, the only serious competitor against Leo’s ATR-1 would be Google’s Gemini 3. They both perform better in different niches, so we decided that it was best for users to have the flexibility to choose which to use in any particular case.
That said, Gemini 3 does tend to “overcorrect” more than ATR-1: to predict what the word should be, based on what it has seen before, rather than what’s actually in the text. It also tends to “correct” (i.e., modernize) orthography in a way that is not faithful to the original transcription. Though very powerful, that’s the main issue with using general-purpose LLMs for transcription, rather than models specially engineered for the task.
We’ll be releasing a new model in the coming months which we’re hoping will smash both out of the water!
Yeah, overcorrection is what I dislike the most about Gemini. It confidently and plausibly makes shit up when the hand is hard, in my experience. Looking forward to the new model!
