Mistral OCR 3

Original Article ↗ Hacker News Discussion ↗

Comparisons with other OCR models

  • Many comments note that recent open-source OCR/VLM systems (PaddleOCR-VL, olmOCR, Chandra, dots.ocr, MinerU, MonkeyOCR, etc.) are strong and often run on smaller, edge-capable models.
  • Several users share external leaderboards where Google’s Gemini models currently rank above Mistral OCR; some say codesota/ocr and ocrarena show Mistral trailing top OSS and proprietary systems.
  • People want head‑to‑head comparisons against these modern baselines, not only against traditional CV OCR engines.

Benchmarks & evaluation transparency

  • Some criticize Mistral’s marketing and benchmark tables as cherry‑picked or unclear, especially around which datasets (“Multilingual,” “Forms,” “Handwritten”) are used.
  • There’s confusion between “win rate” versus “accuracy”: clarification emerges that ~79% refers to how often OCR 3 beats OCR 2, not per‑document correctness.
  • Requests for more failure‑case examples, handwriting benchmarks, and open benchmark data are common.

Performance, accuracy & real‑world use

  • Mixed reports:
    • Some find Mistral OCR 3 inferior to Gemini 3 for complex or historical documents (e.g., 18th‑century cursive, older Scandinavian/Portuguese records), where output is effectively unusable.
    • Others report strong results for math/LaTeX and early experiments replacing MathPix, but Gemini 3 is repeatedly praised for near‑perfect markdown+LaTeX.
  • Concern that a system marketed as “ideal for enterprise” must approach near‑perfect accuracy, especially for scientific and financial documents where small numeric errors are catastrophic.

Hybrid pipelines & “The Way”

  • Several practitioners advocate hybrid setups:
    • Classic OCR (Tesseract, PaddleOCR, RapidOCR, etc.) for boxes/characters, then an LLM/VLM (Mistral, Gemini) for cleanup, structure, and semantic checks.
    • This is seen as safer for high‑accuracy workflows than relying solely on a VLM.

Pricing, API model & developer UX

  • Flat page‑based pricing ($/1k pages) is praised as simpler than token‑based vision billing, though OCR 3 doubling to $2/1k pages annoys some.
  • Others argue per‑character billing would be more transparent, and ask what “a page” size really means.
  • People appreciate a direct OCR API instead of chat UX.
  • Complaints surface about “contact sales” offerings and unresponsive sales teams.

Strategy, ecosystem & deployment

  • Some see Mistral’s focus on OCR/document AI and B2B as smart differentiation from “meme” consumer features; others think they’re being outclassed by US giants.
  • EU regulation and talent attraction are debated: some claim regulation/taxes hinder Mistral; others push back that compliance burden is overstated.
  • Strong demand remains for high‑quality, locally runnable/open models due to privacy and “no cloud for confidential docs,” even as hosted APIs dominate current offerings.