Newer Models, Same Advantage

**DharmaOCR** beat **Mistral OCR4** and **Unlimited-OCR** on **Brazilian Portuguese**, according to a July 16, 2026 Hugging Face Blog post from Dharma-AI authors Erick Lachmann, Gabriel Pimenta de Freitas Cardoso, Francisco de Almeida Rocha Alves, and Victor Gabriel Ferreira Barbosa. The piece, titled Newer Models, Same Advantage, argues the win came from domain specialization and targeted training rather than a newer base architecture. Three months earlier the team had published a paper on DharmaOCR and open-sourced one of the models, with the stated objective of OCR engineered specifically for Brazilian Portuguese.

The training pipeline is described as two stages. The first is supervised fine-tuning on a broad collection of Portuguese-language files drawn from different sources, formats, and levels of complexity. That stage is what aligned the model’s weights for the target language domain before the rest of the pipeline ran.

For engineers building document pipelines that must read Brazilian Portuguese, the useful claim is narrow and testable: a specialized OCR stack beat two named general-purpose OCR systems on that language. Product accuracy on local invoices, forms, and scans often fails when teams ship English-heavy OCR defaults; this post supplies a concrete counter-example and an open model to try on real traffic.

On the competitive side, the framing is explicit. Newer architectures from Mistral OCR4 and Unlimited-OCR did not erase DharmaOCR’s edge on this language slice. The public paper plus an open-sourced model give others a baseline they can re-run, rather than a closed leaderboard claim alone.

What to do next: benchmark DharmaOCR on your own Brazilian Portuguese document mix against the OCR you already ship, and treat architecture freshness as secondary to language and domain fit. Watch for fuller public detail on the second training stage and whether the same ranking holds as Mistral OCR4 and Unlimited-OCR update.