Mistral Small 4: Complete Specifications, Pricing, API Access & Use Cases (2026)

What Is Mistral Small 4?

Mistral Small 4 is Mistral AI’s open-weight multimodal large language model, released on March 16, 2026, featuring a 256K-token context window and hybrid instruct, reasoning, coding, and image-understanding capabilities, with API pricing of $0.15 per 1M input tokens and $0.60 per 1M output tokens as of June 2026.

Mistral AI describes the model as a major Mistral Small family release that unifies capabilities associated with reasoning, multimodal understanding, and agentic coding into one model. The official model card lists 119B total parameters with 6.5B active parameters, making it a mixture-of-experts model designed for efficient inference rather than a dense model of the same total size.

Users usually search for Mistral Small 4 to check whether it is a cost-sensitive alternative to larger frontier models, whether its long context window fits document workflows, and whether it can handle both text and image inputs. Teams comparing it with Mistral Large deployment choices should separate cost, context length, modality support, and reasoning needs rather than treating one model as universally preferable.

What Are Mistral Small 4’s Key Specifications and Pricing?

What Can Mistral Small 4 Do That Makes It Useful in Production?

Mistral Small 4 can support general chat and instruction-following workflows where teams need lower-cost text generation, summarization, extraction, and support automation. Its 256K context window makes it relevant for longer prompts and document-heavy workflows, although long context still requires careful prompt design and retrieval validation.

It can process image inputs for multimodal analysis, including document and visual understanding scenarios. This is useful for extraction and review workflows, but image interpretation should still be checked by humans in high-stakes contexts because visual-language models can misread details.

It supports function calling, structured outputs, and JSON schema mode, which makes it practical for applications that need model outputs to connect with tools, APIs, and typed downstream systems. Understanding GPT-4o mini API patterns can also help teams compare low-cost multimodal and tool-enabled model options across providers.

Mistral Small 4 also includes configurable reasoning effort, allowing developers to choose faster responses for ordinary tasks or deeper reasoning behavior for more complex prompts. This does not eliminate hallucination risk, and outputs still need evaluation for accuracy, latency, and cost in the target workload.

What Are Mistral Small 4’s Supported Modalities?

Where Does Mistral Small 4 Fall Short?

Mistral Small 4 is not a fully specified public system in every operational detail. Official sources confirm context window, pricing, core modalities, features, model IDs, and license, but do not clearly confirm a knowledge cutoff or universal rate limits as of June 2026.

This is a general AI limitation and is not model-specific unless stated by the provider: the model can hallucinate, omit context, misread images, or produce plausible but incorrect outputs. Legal, medical, financial, safety, and compliance decisions should require expert review.

Long-context support does not guarantee perfect long-document reasoning. Even with a 256K-token context window, teams should use retrieval, chunking, citations, evaluation sets, and regression tests for production document workflows. For teams evaluating long-context alternatives such as Gemini 1.5 Pro, context size should be tested against the actual document length, retrieval method, and answer-quality requirements.

The open-weight release may also introduce deployment complexity. Self-hosting or fine-tuning can offer control, but infrastructure requirements, serving stack maturity, quantization choices, and operational monitoring affect latency, throughput, and cost. Mistral’s announcement lists recommended infrastructure examples, which means smaller teams may prefer hosted API access for initial production testing.

What Is Mistral Small 4 Best Used For?

How Does Mistral Small 4 Compare to Mistral Medium 3.5 and Mistral Large 3?

How Do I Access Mistral Small 4?

Mistral Small 4 is available through the Mistral API and AI Studio, with the hosted API alias mistral-small-latest. The official API documentation lists POST /v1/chat/completions, bearer-token authentication, the model field, message-based requests, streaming support, function/tool fields, and structured response formats.

Python Example

from mistralai.client import Mistral
import os
api_key = os.environ["MISTRAL_API_KEY"]
client = Mistral(api_key=api_key)
response = client.chat.complete(
    model="mistral-small-latest",
    messages=[
        {
            "role": "user",
            "content": "Summarize the main production use cases for Mistral Small 4."
        }
    ],
)
print(response.choices[0].message.content)

curl Example

curl https://api.mistral.ai/v1/chat/completions \
  -X POST \
  -H "Authorization: Bearer $MISTRAL_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "mistral-small-latest",
    "messages": [
      {
        "role": "user",
        "content": "Summarize the main production use cases for Mistral Small 4."
      }
    ]
  }'

FAQs

What is the context window of Mistral Small 4?
Mistral Small 4 has a 256K-token context window as of June 2026, according to Mistral AI’s official model card and announcement.

How much does Mistral Small 4 cost?
As of June 2026, Mistral Small 4 API pricing is $0.15 per 1M input tokens and $0.60 per 1M output tokens. Cached prompt tokens are billed at 10% of standard input price when prompt caching applies.

How can developers access Mistral Small 4?
Developers can access Mistral Small 4 through the Mistral API using mistral-small-latest with the /v1/chat/completions endpoint and bearer-token authentication.

What is Mistral Small 4 useful for?
Mistral Small 4 may fit cost-sensitive chat, document understanding, multimodal extraction, coding assistance, agentic workflows, and structured output tasks, provided the deployment includes evaluation, human review for high-stakes use, and hallucination controls.