Mistral AI: Key Models, Tools, and Features

Public assistant, developer agents, platforms for training and deploying, and a portfolio of largely open models: Mistral AI's offerings cover the entire chain. It combines performance metrics, such as responses of up to 1,000 words per second and a LLM with 675 billion parameters, with specialized components for voice, OCR, or moderation. Here’s a structured overview of what is available and what remains proprietary.

Responses at 1,000 words/s and moderation by classifier

Vibe features a fast generation mode, Flash Answers, which relies on updated web resources and AFP news. This mode can reach approximately 1,000 words per second. Mistral leverages the entire AFP archive since 1983. For moderation, Shieldstral, unveiled in August 2026, acts as a security classifier capable of analyzing texts and images based on rules expressed in natural language. The tool returns a risk score intended to filter problematic content. For topics requiring sourced bases, a deep search mode compiles and synthesizes web sources to produce a detailed and documented response.

Open licenses and proprietary scope

Most models are offered in open source or open weight under the Apache 2.0 license, allowing for unrestricted downloading and deployment, including for commercial purposes. However, some advanced models and certain services remain proprietary and are only accessible via Vibe and the API. The model offerings are divided between generalist and specialized, and Mistral mentions the existence of other even more targeted variants. The complete list is available in Mistral's documentation.

Three generalist ranges and specialized components

In the generalist model category, Mistral Large 3, unveiled in December 2025, combines a mixture of experts architecture with 675 billion parameters and offers a context window of 256,000 tokens. This model, capable of handling multiple languages and modalities, is described as the most powerful in the range and suitable for agentic uses. Mistral Small 4, launched in March 2026, integrates reasoning, vision, and code into a single architecture, with the ability to adjust the level of reasoning according to the task at hand. Mistral Medium 3.5 is designed for long tasks, tool invocation, and agentic code; it serves as the reference model for most uses and corresponds to the default model of Vibe. For embedded uses, the Ministral 3 range offers 3B, 8B, and 14B versions suited for resource-limited devices, from smartphones to embedded equipment, while maintaining multimodal and multilingual functions. Specialized models include Codestral for rapid code completion, Devstral 2 for autonomous agents capable of intervening on an entire codebase, Voxtral for voice (transcription, recognition, synthesis, and cloning, used as the engine for Vibe's voice mode), and Mistral OCR for extracting, understanding, and rendering complex documents.

Tools for coding and orchestrating agents

Vibe for code presents itself as a development agent available in terminal interface, extensions for VS Code, JetBrains IDEs, and Zed, as well as background agents. It can read project files, modify code, execute commands, and open pull requests under supervision. For building AI applications and agents, Studio offers a design, deployment, and management platform, with access to models via API, a prompt testing bench, and consumption tracking.

Enterprises, assistant, and integrations: concrete uses

For organizations, Forge offers training, aligning, and evaluating custom models based on Mistral models to adapt them to internal data. The Mistral Compute infrastructure runs these models at scale and is open to laboratories, research teams, and companies, for both training and inference. On the daily usage side, Vibe is accessible on the web, iOS, and Android, with three modes: Chat, Work, and Code. The assistant responds to queries, can analyze documents, generate images, and launch online searches. User-added skills activate automatically based on demand. Scheduled tasks automate monitoring, tracking, or reminders. A Canvas allows for drafting reports, summaries, and presentations, then exporting to Notion, SharePoint, or messaging. The tool generates and edits visuals. A memory feature, via Souvenirs, retains details from conversations to improve relevance, with options for consultation, modification, and deletion. Projects gather reusable files, libraries, and parameters. Connectors link Vibe to business tools and support the MCP protocol. Document analysis covers PDFs and images relying on OCR and multimodal understanding. The entire suite evolves rapidly with regular announcements of new tools, models, and features.