Generalist Models

Mistral AI’s generalist lineup

Mistral AI offers a complete lineup of generalist models, from the most compact to the most powerful, each designed for a different balance of performance, cost, and capability. This diversity isn’t just catalog padding: it reflects usage reality. Writing an email, analyzing a fifty-page contract, and orchestrating an autonomous agent are three tasks that demand neither the same power nor the same budget. Understanding who does what in this lineup lets you choose the right tool for every situation—and that’s exactly what this lesson covers.

Key figures for generalist models

Mistral Medium 3.5 — The frontier model (May 2026)

Mistral Medium 3.5 (announced May 22, 2026, version 26.04, Open tier—open weights under a modified MIT license) is the lineup’s newest frontier model. It succeeds Mistral Medium 3 and sits at the top of the offering for demanding tasks. Three qualities set it apart. It is multimodal: it processes text and images in a single request, which lets you, for example, submit a diagram accompanied by a written question. It is agentic: it was designed to orchestrate agents, meaning it chains tool calls and handles multi-step tasks without constant supervision. Finally, it excels at code, both in generation and analysis.

When quality trumps everything else—strategic analysis, multi-step reasoning, a critical agentic workflow—Medium 3.5 is the model to consider first.

Mistral Small 4 — The hybrid all-rounder (March 2026)

Mistral Small 4 (version 26.03) is likely the model you’ll use most often, and it deserves attention. Its originality lies in its hybrid architecture, which combines three capabilities usually spread across multiple models: instruct mode for answering questions and following instructions, step-by-step reasoning for complex problems, and code generation. The model switches automatically from one mode to another based on the nature of your request—you don’t have to configure anything.

Add to that a context window of 256,000 tokens, roughly 500 pages of text, and you get a tool capable of absorbing an annual report, a contract, or a complete technical manual in a single request. Small 4 is an Open tier model: its weights are available for download, and it is multimodal (text and images). Its built-in reasoning mode now covers needs previously addressed by the Magistral family, now deprecated—we’ll come back to that below.

Mistral Large 3 — The open-weight generalist

Mistral Large 3 (December 2025, version 25.12) is the lineup’s large open-weight generalist model. Multimodal—text, images, documents—it targets complex tasks: long document analysis, multi-step reasoning, elaborated content production. Its decisive advantage lies elsewhere: since its weights are available, you can host it on your own infrastructure. For a company whose data must never leave internal servers—a law firm, a healthcare facility—this makes all the difference. Large 3 remains a safe bet in this scenario; and for self-hosted frontier performance, Mistral Medium 3.5, also open-weight, is now an option.

The Ministral lineup — Compact models

At the other end of the spectrum, the Ministral 3 models are the lineup’s compact versions, optimized for scenarios where response speed and cost take precedence over sophistication. Three sizes are available, all released in December 2025 (version 25.12, Open tier): Ministral 3 14B, the most capable of the compacts, with text and vision; Ministral 3 8B, a good balance between performance and efficiency, also equipped with vision; and Ministral 3 3B, the smallest, designed for deployments on mobile devices or at the edge (edge computing).

Concretely, these models shine where a large model would be overkill: an application that must respond to thousands of simple requests per hour, a service running on a modest server, a prototype to validate without committing budget, or a classification task that doesn’t justify the power of a Large or a Medium. Paying frontier model prices to sort emails would be wasteful—Ministral exists precisely for that.

What about reasoning? The end of Magistral

The Magistral subfamily, formerly dedicated to reasoning, is now deprecated: all its variants appear in the list of removed models, with removals staggered between late 2025 and mid-2026. Reasoning hasn’t disappeared, however—it has been integrated into the generalists. Mistral Small 4 unifies instruct, reasoning, and code in a single model, and Mistral Medium 3.5 takes over for the most demanding reasoning. You no longer need to choose a separate model to reason: reasoning mode is part of the standard lineup.

How to navigate this lineup

To summarize, keep this framework in mind:

• Daily use → Mistral Small 4
• Demanding tasks / agents → Mistral Medium 3.5
• Self-hosting → Mistral Large 3, or Medium 3.5 for frontier (both open-weight)
• Tight budget / high volume → Ministral 3 (8B or 14B)
• Embedded / mobile → Ministral 3 3B

The “Choosing the Right Model” lesson will deepen this selection method with a complete decision tree.

Key takeaways

• Mistral Medium 3.5 (May 2026) is the frontier model: multimodal, agentic, strong at code—and open-weight
• Mistral Small 4 (March 2026) is the hybrid instruct/reasoning/code all-rounder, with 256k context tokens
• Mistral Large 3 is the large open-weight generalist, ideal for self-hosting
• Ministral 3 (3B, 8B, 14B) deliver solid performance in compact formats
• Magistral is deprecated: reasoning now lives directly in the generalists