Mistral Ships OCR 4 With Self-Hosted Document Extraction

Mistral's recent changelog covers two specialised releases rather than a new flagship - Mistral Large 3 remains the general model on our roster.

Mistral OCR 4

Per Mistral’s changelog: bounding boxes, block classification and inline confidence scores alongside extracted text, across 170 languages spanning 10 language groups. It runs in a single container for fully self-hosted deployment.

The confidence scores and bounding boxes are what separate this from generic vision-model text extraction: you get per-block geometry and a machine-readable signal for where the model is unsure, which is what document pipelines actually need to route exceptions to human review.

Leanstral 1.5

Model ID labs-leanstral-1-5, an updated Lean 4 formal proof engineering model with improved supervised fine-tuning mixture quality and extended long-context reasoning. It is scheduled for retirement on 30 September 2026, so treat it as a short-lived research release rather than something to build on.

On infrastructure, Mistral opened a 10 MW facility at Les Ulis in Essonne dedicated to inference, scheduled for Q3 2026.

Timing note: Mistral's changelog does not attach precise dates to every entry, so this item is dated to our publication rather than to a release date we could verify.