A Smaller AI Trained on One Language Just Beat Two Bigger, Newer Models at Reading Brazilian Portuguese
DharmaOCR outscored both Mistral OCR4 and Unlimited-OCR on a Portuguese reading test, and the reason comes down to focus, not size.
Key points
- DharmaOCR scored 0.925 on a Brazilian Portuguese benchmark, compared to 0.798 for Mistral OCR4 and 0.7587 for Unlimited-OCR.
- Both competing models were released after DharmaOCR and backed by larger research teams.
- DharmaOCR was trained in two stages: first on Portuguese documents specifically, then on comparative feedback to cut errors and reduce wasted computing time.
- The gap was most visible on real-world Brazilian documents such as ENEM essays, Brazil's national high school exam.
- The advantage comes from concentration: every part of the model points at one language rather than dozens.
A small, focused AI model just outperformed two newer, better-resourced rivals at reading text in Brazilian Portuguese. The margin was not close.
DharmaOCR, an OCR model (software that reads text from scanned documents and images and converts it into editable words) built specifically for Brazilian Portuguese, scored 0.925 on a dedicated Portuguese benchmark. Mistral OCR4 scored 0.798. Unlimited-OCR scored 0.7587. That is a gap of 13 to 16 percentage points in favour of the smaller, older, more specialised tool.
The researchers shared their findings on Hugging Face, the platform where AI teams publish models and papers.
Why did a more specialised model win?
Specialisation won because every part of the model was pointed at the same target. When an AI is trained across dozens of languages, its capacity spreads thin. When it is trained on one language, every parameter, every internal connection, can concentrate on that language's vocabulary, spelling patterns and document quirks.
DharmaOCR was built in two stages. The first stage trained the model on a wide range of Portuguese-language documents at different complexity levels. The second stage used a technique called Direct Preference Optimization, where the model learned not just what the correct answer was, but which of two competing outputs was better. That second stage cut a common failure mode in AI text tools: the tendency to loop, repeat or produce garbled nonsense under pressure.
The combination made the model both accurate and stable.
The practical difference showed up clearly on ENEM essays, handwritten exam papers from Brazil's national high school test. These documents mix cursive writing with names, slang and cultural references that are specific to Brazil.
Mistral OCR4 read the name Chico Buarque, one of Brazil's most famous musicians and poets, as "Chico Barque." Unlimited-OCR rendered the same name as "chico bique" and turned a Buarque quotation into near-gibberish. DharmaOCR read both correctly.
These are not random slips. A model trained lightly on Brazilian Portuguese will fail at precisely the words that make Brazilian Portuguese distinct. Famous names are not edge cases. Getting them wrong is a sign the model did not spend enough time in this particular linguistic space.
The broader lesson here matters for anyone choosing AI tools for a specific job. A model that does everything tends to do each thing less well than one built for your exact task. That is survivorship bias working in reverse: the headline numbers from a big multilingual model may look impressive, but on your documents, in your language or your industry, a focused tool may simply win.
Takeaway: Before paying for the biggest AI tool available, test the one built for your exact task. Benchmark scores on broad tests do not always predict what happens on your specific documents.
