FinixDoc takes on the financial documents OCR keeps mangling

Optical character recognition typically breaks down on real financial paperwork: insurance claims, medical receipts and bank statements that are dense, inconsistently formatted and full of visual noise a clean invoice never has. Ant Open Source, the open-source arm of InclusionAI, released FinixDoc to target that specific failure mode, training the model on domain-specific text distortions such as homoglyph substitutions and shaping its reinforcement-learning reward around getting text, layout, tables and reading order right together rather than scoring them separately.

In the announcement, the team reports a score of 81.43 on its own FinixDocBench and a 92% success rate specifically on large, dense pages. Because those are self-reported figures from the model's creator, the useful check is that both the dataset and a technical report were released alongside the model, so the claim can be tested rather than taken on faith.