Ministral 3 3B

What is Ministral 3 3B?

Ministral 3 3B is the smallest model in Mistral AI’s Ministral 3 family. It is a compact vision-language model: it can process ordinary text as well as images, but its generated responses are text. The model is intended for applications that need useful language and image-understanding capabilities without the hardware, latency, or operating cost associated with larger models.

Mistral AI released Ministral 3 3B on December 2, 2025, under the Apache 2.0 license. That license and the publication of the model weights make the model suitable for self-hosted and customized deployments, subject to the license and any separate obligations that apply to a particular implementation. It is also available through Mistral AI’s hosted services, where the canonical hosted model identifier is ministral-3b-2512.

Within the Ministral 3 lineup, the 3B model is the efficiency-focused option. The family also includes larger 8B and 14B sizes, as well as base, instruct, and reasoning variants. Those related versions may be preferable when a task needs more reasoning depth or stronger coding performance, but Ministral 3 3B is the more practical starting point when low cost, speed, and deployment flexibility matter most.

Capabilities and supported modalities

Ministral 3 3B accepts text and image inputs and produces text output. Image input means that an application can provide a visual document, screenshot, photograph, chart, or other supported image and ask the model to describe, classify, extract, or reason about its contents. It does not natively generate images, audio, video, speech, music, or embeddings.

• Text input and text generation
• Image understanding
• Function calling and tool use
• Structured outputs
• Document question answering
• Prefix completion
• Batch processing

Structured outputs are useful when the application needs predictable fields, such as extracting an invoice number, classifying a support request, or returning a list of document findings. Function calling allows the model to request an application-defined operation, such as looking up a record or passing extracted information to another system. The model does not perform those external actions by itself; the surrounding application must execute approved functions and return their results.

Mistral’s documentation also identifies chat completions, streaming, and prompt-caching mechanisms in the hosted environment. These are useful delivery and inference features, but they should not be confused with additional output modalities. The model remains text-output only.

Context window and output limits

The documented context window is 256,000 tokens. A token is a unit of text used by the model; it may represent a whole word, part of a word, punctuation, or another small piece of content. A 256k-token context allows an application to provide substantially more source material than a conventional short-context model, including long documents, multiple files, or extended conversation history.

The context limit covers the prompt and generated response together. In practical terms, a very large input leaves less room for the answer. Mistral AI does not publish a separate exact maximum-output-token value for this model in the reviewed model documentation, so the maximum response length should not be represented as a known fixed number.

Long context does not guarantee that every detail in a very large input will receive equal attention. For document workflows, results should still be checked, especially when the answer depends on a small detail buried in a large collection of material.

Pricing and deployment options

Mistral AI’s current standard hosted pricing lists Ministral 3 3B at $0.10 per million input tokens and $0.10 per million output tokens. Cached input is listed at $0.01 per million tokens. These are usage-based API prices rather than a recurring consumer subscription price.

The open-weight release provides another deployment path. Organizations can run the model locally or on their own infrastructure when control over data location, latency, or operating environment is more important than using a managed endpoint. The actual hardware requirements, throughput, and total self-hosting cost depend on the inference stack, quantization, workload, and deployment configuration; the supplied documentation does not establish one universal hardware profile.

Reasoning, coding, and tool use

Ministral 3 3B supports ordinary instruction-following, document question answering, structured extraction, and lightweight agent workflows. Its editorial reasoning score is 6 out of 10, but that score is a comparative assessment rather than a Mistral AI benchmark or published vendor rating. The model should be viewed as capable of practical, bounded reasoning rather than as a frontier system for difficult multi-step analysis.

The same distinction applies to coding. The editorial coding score is 5 out of 10. Ministral 3 3B can assist with code generation, classification, extraction, and task routing, especially when a surrounding application provides clear instructions and tools. It is not the strongest choice for large software projects, complex debugging, or situations where high coding reliability is more important than speed and cost.

Function calling gives the model a way to participate in tool-based workflows. For example, it could read a user’s request, select a weather or inventory function, provide structured arguments, and then turn the returned result into a response. This makes it suitable for lightweight agents and business automation, provided that the application validates arguments, controls permissions, and handles errors.

Main strengths and trade-offs

• Low operating cost: The listed input and output rates are low enough to make high-volume classification, extraction, and routing workloads practical.
• Fast response potential: Its small size is suited to low-latency applications. The editorial speed score is 9 out of 10, which is an estimate rather than a provider-published performance guarantee.
• Deployment flexibility: Apache 2.0 licensing and open weights support local, private, and customized deployments, while the hosted API avoids the operational work of running infrastructure.
• Vision-language input: The model can work with images as well as text, enabling image-aware assistants and document workflows without requiring a separate text-only model for every request.
• Long context: The 256k-token context window supports large document and multi-file use cases.
• Application integration: Function calling, structured outputs, batching, and streaming support practical production workflows.

The central trade-off is capability versus efficiency. A 3B-parameter model is designed to be economical and responsive, not to maximize reasoning depth. Larger models in the Ministral 3 family or Mistral AI’s broader catalog may be more appropriate for difficult mathematics, complex planning, demanding software engineering, or tasks requiring more robust instruction following.

Where Ministral 3 3B works well

Ministral 3 3B is a good fit when a workload combines moderate language capability with strict cost, latency, privacy, or hardware constraints. Practical examples include:

• Local assistants that answer questions without sending every prompt to a third-party hosted service
• Document question answering over manuals, reports, contracts, or internal procedures
• Image-aware classification, such as sorting incoming visual documents or identifying categories in screenshots
• Structured extraction from text and images into application-defined fields
• Lightweight agents that call a small set of approved business functions
• Task routing, intent classification, and triage before a request is sent to a larger model
• Edge-device applications where network latency or continuous cloud access is undesirable
• Batch processing of large collections of documents or records

It is especially attractive when the application can constrain the task clearly. A fixed extraction schema, a small tool set, or a defined classification taxonomy helps the model deliver useful results while reducing the need for open-ended reasoning.

When to choose another model

Choose a larger model when the main requirement is difficult reasoning rather than economical inference. Larger Ministral 3 variants are more suitable when additional capacity is needed, although the supplied research does not provide a direct benchmark comparison between the sizes. Mistral AI’s larger general-purpose models may also be preferable for demanding coding, complicated planning, or higher-reliability instruction following.

A different model category is appropriate when the required output is not text. Ministral 3 3B does not generate images, audio, video, or speech, and it does not provide embeddings. An application requiring any of those outputs needs a dedicated model or service for that function.

The model is also not an ideal choice when a documented knowledge-cutoff date is essential. Mistral AI has not published a direct knowledge-cutoff date for this exact model in the reviewed first-party documentation. Current or rapidly changing information therefore requires retrieval, external grounding, or another verified data source rather than reliance on the model’s internal knowledge alone.

Bottom line

Ministral 3 3B is a practical compact model for text generation, image understanding, structured extraction, document question answering, and lightweight tool-driven applications. Its combination of open weights, Apache 2.0 licensing, a 256k-token context window, and low hosted pricing makes it particularly relevant to edge, local, privacy-sensitive, and high-volume workloads.

Its limitations are equally important: it is text-output only, has no documented exact maximum output value, lacks a published knowledge-cutoff date, and should not be expected to match larger models on difficult reasoning or complex coding. The best reason to choose it is not maximum capability in every task, but a favorable balance of speed, cost, deployment control, and multimodal input for focused applications.