Mistral's New OCR Includes Catalan, but It's Not Open Source
The tool recognises text, detects its position on the page, and identifies titles, explicitly listing Catalan among supported languages. The company promises greater speed and lower cost, though this cannot be verified.
Mistral has launched a new OCR that promises to be particularly noteworthy. It not only recognises text but also tells you its position on the page with bounding boxes, and detects titles. Catalan is explicitly listed as one of the supported languages, something not always common with tools of this kind.
An OCR (optical character recognition) converts text images (scanned documents, photographs, PDFs) into editable and searchable text. The ability to detect position and structure, not just letters, is what allows tables, forms, and complex documents to be reconstructed without losing their format. For archives, administrations, and media working with ample documentation in Catalan, explicit language support is a practical advantage.
The creators claim that it achieves the same accuracy as current providers, but at seventeen times the speed and at eight times lower cost. However, these are assertions from Mistral that we can't verify: we cannot test it, and it's not open source.
It's worth setting the context. Mistral has been known for releasing several open-source products, open-weight models like Mistral Small, Mistral Large, and smaller models of 3B, 8B, and 14B parameters, Mistral NeMo, Mixtral 8x7B, and Mixtral 8x22B. This product, the OCR 4, is not one of them. The distinction is important: with an open-weight model, you can audit it, deploy it independently, and not be reliant on a provider. With a closed model, you cannot. Despite this, the press release states that it can be run as local software, on-premises, in a single container, which at least allows processing documentation without sending it to a third party.
For those digitising document collections in Catalan, like libraries, archives, and administrations, the combination of recognising structure and having the capability to operate locally is precisely what's needed to avoid relying on an external service and handing over sensitive documents. However, it remains to be seen if this can be verified once the tool is thoroughly tested by someone.