Mistral OCR 4(mistral.ai)
472 points by meetpateltech 22 hours ago | 124 comments
tl;dr: Mistral released OCR 4, a document extraction model that returns bounding boxes, typed block classifications (tables, equations, signatures, etc.), and per-word confidence scores across 170 languages, deployable in a single self-hosted container. It claims top scores on OlmOCRBench (85.20) and 72% win rates in human preference tests against competitors, though Mistral notes benchmark scoring artifacts inflate apparent errors on math and multi-column docs. Pricing is $4/1k pages via API ($2 batch), $5/1k for Document AI, available through Mistral Studio, AWS SageMaker, and Microsoft Foundry.
HN Discussion:
  • Skepticism about Mistral's benchmark claims given past versions underperformed external benchmarks
  • Criticism of misleading presentation like truncated y-axes and selective benchmark reporting
  • Questioning the price increase and lack of clear differentiation from the previous v3 model
  • ~Mixed real-world testing results showing accuracy issues with certain languages/scripts compared to competitors
  • Curiosity about missing competitor comparisons (e.g., Claude) in the benchmarks