Small 4, Large 3, and Ministral 3: The General-Purpose Family

Updated November 2, 2026. Mistral Large 3 (mistral-large-2512) has been deprecated as of October 31, 2026, with scheduled removal on April 30, 2027, replaced by Mistral Large 4 (Lesson 10). Mistral Small 4 and the Ministral 3 series remain in general availability.

A single frontier model doesn’t make a lineup. Around Medium 3.5, Mistral aligns a full general-purpose family that covers the cost-to-capacity spectrum, from the top-tier self-hostable model to the three-billion-parameter model built for embedded systems. Understanding this family means understanding how to choose: most of the billing or latency errors we see from learners come from using a model oversized for the task. Here is the lineup as the official documentation presents it as of July 28, 2026.

The Official Snapshot

One detail runs through the entire table: each of these models bears the “OPEN” label in the documentation. The open-weight strategy remains a Mistral hallmark, from the smallest Ministral to the frontier model—which, for organizations subject to sovereignty constraints, changes the very nature of the choice: it’s not a question of whether self-hosting is possible, but rather at which tier of the lineup it makes sense.

Small 4, or the End of Standalone Reasoning

The official description of Small 4 deserves a closer look, because it encapsulates an entire market shift in one word: it is a “hybrid” model, unifying instruction, reasoning, and code in a single efficient architecture. Until spring, the Mistral landscape separated these capabilities—one model for conversation, a Magistral model for reasoning. Small 4 merges the two. This is the same trend among other publishers in 2026: reasoning ceases to be a separate model to become an integrated capability, which the model engages when the task demands it. You already know the direct consequence if you read Lesson 1: this is precisely what explains the deprecation of Magistral, which Lesson 8 covers in detail. For now, keep the practical implication in mind: for everyday use with an occasional need for reasoning, a single model now suffices, where yesterday you had to orchestrate two.

Ministral 3: Edge in Three Sizes

At the other end of the spectrum, the Ministral 3 series offers the same formula in three sizes—3B, 8B, and 14B—with a notable constant: all handle text and vision. A three-billion-parameter model capable of analyzing an image is what makes fully local scenarios credible: document reading on an isolated workstation, image analysis on industrial equipment, offline embedded assistant. For sovereign use cases—local deployment, sensitive data, no cloud dependency—this is the natural path, consistent with what our Mistral deployment courses already describe. The size choice follows the expected logic: 14B brings the series’ maximum power, 8B offers balance, and 3B provides efficiency for the most constrained environments.

How to Choose, in Practice

The selection logic comes down to two questions. First question: are you self-hosting? If yes, Large 3 is the choice for maximum capacity—it is the top-tier open-weight multimodal general-purpose model—and Ministral 3 takes over as soon as hardware constraints or on-site requirements demand it. If you use the hosted API, second question: does the task justify the frontier model? Medium 3.5 is the default choice for agents, code, and demanding multimodal tasks; for everything else—and that’s the majority of traffic in a real-world application—Small 4 offers the best capacity-to-cost ratio, reasoning included. A useful exercise: take three typical queries from your own usage and ask yourself, for each, whether it really needs Medium 3.5 or if Small 4 would suffice. In our experience, the honest answer is “Small 4” more often than you might think.

Official Source: docs.mistral.ai — Models Overview

Key Takeaways

• Small 4 unifies instruction, reasoning, and code in a single model
• Large 3 is the open-weight general-purpose model; Ministral 3 covers the edge in 3B/8B/14B with vision
• The entire general-purpose family is open-weight
• The choice is between hosted API (Medium/Small) and self-hosting (Large/Ministral)
• Reserve Medium 3.5 for tasks that justify it: Small 4 covers everyday use