Mistral OCR 4 makes self-hosted document AI good enough to actually use

Document ingestion has been the ugly, underfunded stage of every RAG pipeline I've built. Mistral just made it a lot less ugly, and you can run it in your own VPC.

On June 23, Mistral released OCR 4, a document-intelligence model that returns bounding boxes, block classification, and inline confidence scores across 170 languages, and runs fully self-hosted in a single container (Mistral AI). It tops OlmOCRBench at 85.20 with a 72% average human-preference win rate over rival OCR systems, priced at $4 per 1,000 pages, or $2 via the Batch API. The pitch is aimed squarely at regulated teams who can't send documents to a cloud API, which is a real and large group.

The part that matters to builders isn't the benchmark, it's the structure. Plain OCR gives you a wall of text and you reconstruct layout yourself, badly. Bounding boxes plus block classification plus per-block confidence means you can chunk on actual document structure (this is a table, this is a heading, this is a footnote) and you can route low-confidence blocks to a human or a second pass instead of silently poisoning your index. I've shipped pgvector RAG where 80% of the retrieval quality problems traced back to garbage chunking on top of garbage OCR. Fix the ingestion layer and a lot of "the model is dumb" complaints disappear.

The self-hosted-in-one-container detail is the unlock for anyone in finance, health, or legal. I spent years at a fintech platform where "can we send this to an external API" was a multi-week conversation that usually ended in no. An on-prem model that clears 85 on OlmOCRBench means the answer can finally be yes without a vendor risk review.

What to do: if you have a private-corpus pipeline, benchmark OCR 4 against whatever you're running now on your own documents, not theirs. Measure structure recovery, not just character accuracy. And if you're cloud-OK, the $2 Batch price is cheap enough that the real question is whether the structured output saves you enough downstream engineering to justify the swap. For me it probably does.