Mistral OCR 3: Independent Benchmark Results
We ran the full OmniDocBench (1,355 images) and OCRBench v2 (7,400 English samples) benchmarks ourselves. Here's what we found.
Quick Install
pip install mistralaiAPI access via api.mistral.ai | Python SDK available | Pay-per-use pricing
Documentation
Tutorial
Learning-oriented. Extract text from your first PDF with Mistral OCR 3 in 5 minutes.
Start here →How-To Guides
Problem-oriented. Batch processing, invoice extraction, table parsing.
Solve problems →Reference
Information-oriented. API endpoints, parameters, response formats.
Look it up →Explanation
Understanding-oriented. Architecture, when to use it vs alternatives.
Understand →Independently Verified Benchmarks
CodeSOTA ran two full benchmark suites: OmniDocBench (Dec 19, 2025) and OCRBench v2 (Dec 21, 2025). We processed 8,755 images total through Mistral's OCR API and computed metrics using official evaluation tools.
OmniDocBench Results (Verified)
OmniDocBench is a comprehensive document parsing benchmark with 1,355 pages across 9 document types. The composite score formula: ((1-TextEditDist)*100 + TableTEDS + FormulaCDM) / 3
| Metric | Mistral OCR 3 | GPT-4o | PaddleOCR-VL |
|---|---|---|---|
| Composite Score | 79.75verified | ~85 | 92.86 |
| Text Edit Distance | 0.099verified | 0.02 | 0.03 |
| Table TEDS | 70.9%verified | - | 93.5% |
| Table Structure TEDS | 75.3%verified | - | - |
| Formula (Edit Distance) | 0.218verified | - | - |
| Reading Order | 91.6%verified | - | - |
Lower is better for Edit Distance metrics. CodeSOTA verification date: December 19, 2025.
OCRBench v2 Results (Verified)
Important: Mistral OCR is a pure OCR model - it extracts text/markdown from images. OCRBench v2 tests 21 task types including VQA, chart parsing, and structured extraction, which require a Vision-Language Model (VLM). Mistral OCR excels at text extraction tasks but cannot answer questions about images.
| Task Type | Score | Samples | Notes |
|---|---|---|---|
| Full-Page OCR | 79.1% | 200 | Core OCR strength |
| APP Agent | 69.5% | 300 | UI text extraction |
| Cognition VQA | 58.3% | 800 | Some VQA works via text |
| Math QA | 56.2% | 300 | |
| Document Parsing | 55.2% | 400 | Structured output |
| Science QA | 44.0% | 300 | |
| Text Recognition | 32.5% | 800 | |
| Table Parsing | 4.4% | 400 | Weak on OCRBench format |
| Chart/Spotting/Grounding | 0% | various | Requires VLM, not pure OCR |
CodeSOTA verification date: December 21, 2025. Full results at OCRBench v2 Leaderboard.
Context: Top VLMs score 55-62% on OCRBench v2 (Seed1.6-Vision: 62.2%, GPT-4o: 47.6%). Mistral OCR's 25.2% reflects that it's an OCR tool, not a VLM - comparing these is like comparing a scanner to a research assistant.
Performance by Document Type
Mistral OCR 3 performs best on academic papers and exam papers, struggles with newspapers:
| Document Type | Text Accuracy | Table TEDS |
|---|---|---|
| Academic Literature | 97.9% | 83.0% |
| Exam Papers | 92.8% | 88.0% |
| Books | 93.9% | 82.7% |
| Research Reports | 95.8% | 82.0% |
| Magazines | 97.9% | 71.0% |
| PPT Slides | 95.7% | 72.6% |
| Newspapers | 67.0% | 58.3% |
Performance by Language
Text accuracy
Text accuracy
Text accuracy
Pricing
Real-time processing via API
50% discount for async processing
Our full benchmark run cost $2.71 for 1,355 images using the standard API.
Code Example
from mistralai import Mistral
client = Mistral(api_key="your-api-key")
# Load document
with open("document.pdf", "rb") as f:
doc_data = base64.b64encode(f.read()).decode()
# OCR with Mistral OCR 3
response = client.ocr.process(
model="mistral-ocr-2512",
document={"type": "pdf", "data": doc_data}
)
# Output is markdown with HTML tables
print(response.content)When to Use Mistral OCR 3
- Academic papers (97.9% accuracy)
- Exam papers (92.8% + 88% tables)
- Research reports and books
- Cost-sensitive high-volume OCR
- English text extraction
- Newspapers (67% - avoid)
- Complex multi-column layouts
- Chinese text (86% vs 94% English)
- Table recognition vs PaddleOCR
Verdict
Mistral OCR 3 is a dedicated OCR tool, not a VLM. It excels at text extraction (79.1% full-page OCR, 90.1% text accuracy) but cannot answer questions about images or parse charts.
For document parsing, it's behind PaddleOCR-VL (92.86) but faster and cheaper. For VLM tasks, use Qwen3-Omni or GPT-4o instead.
Best use case: High-volume text extraction from PDFs/documents where you need markdown output, not image understanding.
Model ID: mistral-ocr-2512
API: docs.mistral.ai
Verified: OmniDocBench Dec 19, OCRBench v2 Dec 21, 2025 by CodeSOTA