Specialist Models

Models Designed for Specific Tasks

The general-purpose models we just discussed can do almost anything—but “almost anything” isn’t “everything at the highest level.” Alongside this range, Mistral AI develops specialized models, each optimized for a specific domain: code, voice, documents, and content safety. In their specialty, these specialists outperform generalists, just as a sworn translator outperforms a good polyglot on a legal deed. Knowing this lineup of specialists lets you use the right tool at the right time.

Codestral — The Code Specialist

Codestral (August 2025, version 25.08, tier Premier) is the model dedicated to code generation and completion, trained specifically on programming data. Its native environment is the developer editor: it intelligently predicts the next lines of code as you type, writes complete functions from a simple description, improves the quality and readability of existing code (refactoring), and can analyze and document complex code submitted to it. A developer typing the start of a sort function sees Codestral suggest the following lines before even thinking them—that efficiency gain explains its adoption. Codestral is also integrated into integrated development environments (IDEs) like Visual Studio Code or JetBrains via dedicated extensions.

Its variant, Codestral Embed (May 2025, tier Premier), plays a different role: it transforms code into vector representations, enabling semantic search across codebases. Finding similar functions, detecting duplicated code, powering a RAG system on code—these use cases all rely on this building block.

Devstral 2 — The Code Agent

Devstral 2 (December 2025, version 25.12, tier Open) goes further than Codestral. It isn’t just a model that generates code: it’s an agent capable of handling complete software engineering tasks. While Codestral completes the current line, Devstral 2 navigates a multi-file project, understands the application architecture, identifies and fixes bugs, implements features by modifying multiple files at once, and writes the unit and integration tests alongside it. The difference between the two is like that between an assistant who completes your sentences and a developer colleague who understands your project. Devstral 2 replaces the previous generation, Devstral 1.x.

Voxtral — Voice in Both Directions

Voxtral TTS (March 2026, version 26.03) is Mistral’s text-to-speech model. It converts text into natural speech with studio quality: expressive voices, accurate intonation, and fine control over speed, tone, and emotion. It is multilingual—French is of course covered—and offers zero-shot voice cloning: from a short audio sample, it reproduces a voice’s timbre. Imagine a company producing its training content in multiple languages with a consistent brand voice, without a recording studio: Voxtral TTS makes this type of use case accessible.

For the reverse direction—voice to text—Mistral offers Voxtral Mini Transcribe Realtime (February 2026, version 26.02), a real-time transcription model adapted for continuous audio streams: voice assistants, live captioning, call centers. For the complete list of available transcription variants, consult the official models page.

OCR 4 — The Document Expert

OCR 4 (June 23, 2026, version 4.0, tier Premier) is Mistral’s optical character recognition service, replacing the previous generation OCR 3. Its mission: transform scanned documents, PDFs, or photos of documents into usable structured text. It doesn’t just extract characters—it understands the layout. Tables, forms, and invoices are restored in their structure; headings, paragraphs, tables, and lists are identified as distinct structural blocks; and each text block is precisely localized on the page thanks to paragraph-level bounding boxes. It all works on multilingual documents, with high accuracy even when scan quality is poor.

This model is particularly valuable for organizations that process a high volume of documents: an accounting firm digitizing supplier invoices, a legal department indexing contracts, a government agency digitizing its files.

Magistral — A Deprecated Family

If you still encounter Magistral in an old project or tutorial, migrate to Small 4 or Medium 3.5.

Mistral Moderation 2 — The Guardian

Mistral Moderation 2 (March 2026, version 26.03, tier Premier) is a model dedicated to content safety. Its role is to monitor what goes in and out of an AI pipeline: it detects inappropriate content (violence, hate speech, sexual content), identifies attempts to manipulate models—known as jailbreaking—and classifies conversations by risk level. Take a customer service chatbot open to the public: without a moderation layer, nothing prevents a malicious user from hijacking the tool or making it produce problematic content in your brand’s name. Moderation 2 acts as this additional security layer, and for this reason, it is essential for any AI application deployed in production and accessible to the public.

Leanstral 1.5 — The Formal Mathematician

The newest addition to the lineup, Leanstral 1.5 (July 2, 2026) is an open source foundation model specialized in formal proofs in Lean 4, a mathematical proof language. It targets mathematics researchers and specialists in formal software verification. You will probably never use it daily, but its existence illustrates Mistral’s strategy well: cover very diverse AI applications, down to the most specialized.

Key Takeaways

• Mistral offers specialist models that outperform generalists in their domain
• Codestral (completion) and Devstral 2 (software agent) cover all development needs
• Voxtral handles voice in both directions: text to speech and real-time speech to text
• OCR 4 transforms documents into structured data, with precise block localization
• Moderation 2 is indispensable for the safety of AI applications in production
• Magistral is deprecated: reasoning moves through Small 4 and Medium 3.5