What Is Mistral? Models, Hosted Catalogs and Open Weights
Mistral is an AI company and model brand offering hosted interfaces and open weights. This article distinguishes general models from OCR services and explains third-party catalog entries and model-specific licenses.
Meet this model
Mistral is a model brand developed by Mistral AI and is also commonly used to refer to its hosted services. It covers general language and multimodal models, with a separate OCR service for document processing. Developers can call hosted interfaces or self-host releases with explicitly published weights. These are distinct from a chat app. Catalogs may also list third-party models, so the platform name does not establish a model’s developer or license.
How general models differ from dedicated OCR
The official October 20, 2026 catalog lists Mistral Large 4 and Small 4 alongside OCR and Voxtral speech models. The Large 4 page marks it as a public preview and describes a multimodal mixture-of-experts model. A general model can answer questions about the material, while a dedicated OCR interface returns document structure. Their output formats need to be handled separately.
For a knowledge base built from scanned PDFs, first use OCR to extract page text and structure, then have a general model answer from that extraction. The OCR interface supports Markdown or HTML tables, but the rows and columns still need checking against the original. Successful extraction and grounded answers are different outcomes; one successful API call does not establish correctness across the document.
Choose an open-weight release by its model card
One specific self-hosting example is Mistral Small 3.1 24B Instruct. Its official card describes image understanding, a 128K context and an Apache 2.0 license. Those specifications belong to that earlier release, not every Mistral model, and its local running conditions should not be transferred to the Large line.
The official licensing help page says most open models use Apache 2.0, while some use a modified MIT license with additional commercial licensing or platform-use conditions for companies whose monthly revenue exceeds US$20 million. Read the target card and actual license. The umbrella description “Mistral open model” does not specify all usage rights.
A hosted catalog does not contain only in-house models
The catalog also lists Z.ai GLM 5.3 and explicitly identifies it as third-party. When assessing a result, record the model developer and full identifier rather than merely saying it ran on the Mistral platform. Unified access does not give every model identical inputs, update schedules or licensing.
For selection, use the same PDF with merged table headers to compare extracted structure and document-grounded answers before deciding whether to combine OCR with a general model. This article does not test the Mistral API or local weights. A public-preview label is not a stability guarantee, and provider benchmark rankings do not establish results on Chinese business tasks.
Checked sources
These sources support the concepts and methods described here. Examples and test scope are labeled separately; check the version of the tools you use.