Chat

Model developer guide

Mistral AI

7 website model guides

Match the size and complexity of the model to a specific job. Mistral’s catalog spans compact text-and-vision models and larger coding-oriented choices. Start with an output you can validate, then distinguish the hosted API, open weights and the narrower controls available in your app session.

ReasoningVisionCoding

Practical guidance

Where should I start?

Editorial starting points, not a performance ranking. Check current app access before choosing a model.

Plan a complex code change

Evaluate Medium 3.5 when code changes depend on multiple interfaces or constraints. Mistral positions the model for agentic and coding work and documents adjustable reasoning. Keep the first deliverable a plan or review that a developer can check.

Tradeoff: Additional reasoning can increase token use and latency. A model positioned for agents does not automatically have repository access or permission to modify files in EZ Ai Assist.

Mistral Medium 3.5 guide

Extract or classify a bounded input

Consider Ministral 3 8B for a constrained document or text operation with a clear answer format. Its official card describes text and vision capabilities, a 256K context and support for structured outputs and function calling.

Tradeoff: Evaluate difficult examples before optimizing for a compact model. Do not assume Medium’s reasoning-effort choices are supported by a different Ministral identifier.

Ministral 3 8B guide

Technical details

What does the provider document?

Official documentation describes the named model or gateway—not every feature available in this app.

Published context is not a generated-answer limit

Medium 3.5 and Ministral 3 8B both publish 256k context on their model cards. A model-specific maximum output was not established from those cards in this review. Do not convert the context figure into a promised answer length, upload allowance or complete-recall guarantee.

Reasoning changes the returned content shape

For mistral-medium-3-5, Mistral documents reasoning_effort high and none. High returns thinking chunks before the final answer; none performs minimal internal reasoning without a visible thinking chunk. The API response can therefore be a list of chunks rather than a plain string, which matters to parsing and streaming.

Conversation history and service features are distinct

Mistral recommends preserving the complete assistant message, including thinking chunks, in multi-turn reasoning workflows. The model cards separately link Chat Completions, structured outputs, function calling and broader agent services. A model’s support for these interfaces does not install tools or configure an application’s history handling.

Complete website coverage

Explore all 7 model guides

Every existing catalog entry for this provider is linked below. Status labels come from the individual website guides; catalog coverage is not a guarantee of current app access. Historical and snapshot pages remain references, not extra active models.

Practical guidance

From a brief to a useful result

Original prompt examples, not generated-output demonstrations. They have not been generation-tested.

Workflow 1

Extract a supplier comparison without inventing cells

Paste a small set of supplier descriptions with labels. If your app supports image input, verify that a photographed table is legible before using it; text excerpts make the first evaluation easier.

From the supplied supplier descriptions, extract supplier name, stated lead time, minimum order, material and any explicit exclusions. Return one row per supplier and use unknown for unstated values. Include the source label beside each populated field. Do not infer a delivery promise from a marketing phrase or convert units without showing the original value.

Check the result: Spot-check every populated cell against the source and review unknown values manually. Reject confident substitutions for missing lead times or order quantities.

Use the Ministral 3 8B guide

Workflow 2

Break a refactor into safe checkpoints

Supply the current interface, callers and migration constraints. Ask for a sequence that preserves behavior before requesting implementation.

Plan a refactor of the interface described below while preserving existing callers. Identify the compatibility boundary, the smallest first change and the regression test required at each checkpoint. Separate optional cleanup from necessary work. Explain how to stop or roll back if an assumption fails, and do not claim to have inspected files outside the supplied examples.

Check the result: Confirm each checkpoint is independently reviewable and that the plan does not combine a behavior change with unrelated cleanup. Have a developer verify the proposed tests.

Use the Medium 3.5 guide

Technical details

What changes inside EZ Ai Assist?

Official Mistral model cards and reasoning documentation were checked for the provider claims here. The website’s existing model references are not a live app inventory. EZ Ai Assist reasoning settings, attachment processing, endpoint and forwarded parameters were not independently inspected.

Website catalog

What does the full model list represent?

The six existing Mistral and Ministral guides retain their names and URLs. They are model references rather than a promise that every version or service endpoint is enabled for every EZ Ai Assist account.

Unverified app interface

Which user controls are established?

This review did not inspect a Mistral-specific reasoning selector, attachment limit, JSON-schema control or agent-tool panel. The practical examples can begin as ordinary text tasks; use richer inputs only when the current app session supports them.

Unverified upstream contract

Which API details need app verification?

The actual model alias, Chat Completions versus Conversations route, reasoning_effort, history preservation and response-format fields were not captured. Neither the direct model card nor the existence of open weights proves those settings are forwarded.

Before you rely on a result

Limits, costs and tradeoffs

  • A smaller model may omit exceptions, while a larger model may still invent an answer. Include failure cases and evaluate the output against original evidence rather than model size.
  • Open-weight availability does not mean every model shares a license or that local serving reproduces the hosted API. Medium 3.5’s card specifies Modified MIT; review the exact model terms for other deployments.
  • Hosted input/output rates, self-hosting expense and EZ Ai Assist subscription charges are separate. This overview does not infer a flat per-document charge or claim that caching will always apply.

Provider references and picker estimates are not EZ Ai Assist customer charges. Read the website pricing explanation.

Compare by task

Other paths to explore

Cohere

Compare evidence-grounded extraction and document answers with models organized around enterprise text workflows.

Anthropic

Evaluate a second family for a staged code review or refactor plan.

Groq

Compare a hosted compact-model workflow while distinguishing the serving platform from the model developer.

Common questions

Mistral AI, without the assumptions

Does Ministral mean a lower-quality answer for every task?

No fixed ranking is established here. A compact model may be sufficient for a narrowly defined task, but it needs testing on ambiguous and difficult inputs as well as easy examples.

Does 256K context mean it can write 256K tokens?

No. Context and maximum output are different limits. The reviewed cards publish context but did not establish a separate maximum-output figure for these featured models.

Can I use high reasoning with every Mistral model?

Do not assume so. The cited reasoning guide explicitly covers Medium 3.5 and a Small alias. Settings need model-specific confirmation and must also be exposed or forwarded by the app.

New to the terminology?

Official documentation

Sources & review notes

Highlighted provider contracts reviewed . Individual model guides carry their own review dates. Sources can change; these pages do not perform live health checks.

Official Mistral model cards and reasoning documentation were checked for the provider claims here. The website’s existing model references are not a live app inventory. EZ Ai Assist reasoning settings, attachment processing, endpoint and forwarded parameters were not independently inspected.