Pixtral Large is now available in Amazon Bedrock

Today, we are excited to announce that Mistral AI’s Pixtral Large foundation model (FM) is generally available in Amazon Bedrock. With this launch, you can now access Mistral’s frontier-class multimodal model to build, experiment, and responsibly scale your generative AI ideas on AWS. AWS is the first major cloud provider to deliver Pixtral Large as a fully managed, serverless model.

In this post, we discuss the features of Pixtral Large and its possible use cases.

Overview of Pixtral Large

Pixtral Large is an advanced multimodal model developed by Mistral AI, featuring 124 billion parameters. This model combines a powerful 123-billion-parameter multimodal decoder with a specialized 1-billion-parameter vision encoder. It can seamlessly handle complex visual and textual tasks while retaining the exceptional language-processing capabilities of its predecessor, Mistral Large 2.

A distinguishing feature of Pixtral Large is its expansive context window of 128,000 tokens, enabling it to simultaneously process multiple images alongside extensive textual data. This capability makes it particularly effective in analyzing documents, detailed charts, graphs, and natural images, accommodating a broad range of practical applications.

The following are key capabilities of Pixtral Large:

• Multilingual Text Analysis – Pixtral Large accurately interprets and extracts written information across multiple languages from images and documents. This is particularly beneficial for tasks like automatically processing receipts or invoices, where it can perform calculations and context-aware evaluations, streamlining processes such as expense tracking or financial analysis.
• Chart and data visualization interpretation – The model demonstrates exceptional proficiency in understanding complex visual data representations. It can effortlessly identify trends, anomalies, and key data points within graphical visualizations. For instance, Pixtral Large is highly effective at spotting irregularities or insightful trends within training loss curves or performance metrics, enhancing the accuracy of data-driven decision-making.
• General visual analysis and contextual understanding – Pixtral Large is adept at analyzing general visual data, including screenshots and photographs, extracting nuanced insights, and responding effectively to queries based on image content. This capability significantly broadens its usability, allowing it to support varied scenarios—from explaining visual contexts in presentations to automating content moderation and contextual image retrieval.

Additional model details include:

• Pixtral Large is available in the eu-north-1 and us-west-2 AWS Regions
• Cross-Region inference is available for the following Regions:
  • us-east-2
  • us-west-2
  • us-east-1
  • eu-west-1
  • eu-west-3
  • eu-north-1
  • eu-central-1
• Model ID: mistral.pixtral-large-2502-v1:0
• Context window: 128,000

Get started with Pixtral Large in Amazon Bedrock

If you’re new to using Mistral AI models, you can request model access on the Amazon Bedrock console. For more information, see Access Amazon Bedrock foundation models.

To test Pixtral Large on the Amazon Bedrock console, choose Text or Chat under Playgrounds in the navigation pane. Then, choose Select model and choose Mistral as the category and Pixtral Large as the model.

By choosing View API, you can also access the model using code examples in the AWS Command Line Interface (AWS CLI) and AWS SDKs. You can use a model ID such as mistral.mistral-large-2407-v1:0, as shown in the following code:

$ aws bedrock-runtime invoke-model  
--model-id mistral.pixtral-large-2502-v1:0 
--body "{"prompt":"<s>[INST] this is where you place your input text [/INST]", "max_tokens":200, "temperature":0.5, "top_p":0.9, "top_k":50}"  
--cli-binary-format raw-in-base64-out 
--region eu-north-1  
invoke-model-output.txt

In the following sections, we dive into the capabilities of Pixtral Large.

Pixtral Large use cases

In this section, we provide example use cases of Pixtral Large using sample prompts. Because Pixtral Large is built on Mistral Large 2, it includes a native JSON output mode. This feature enables developers to receive the model’s responses in a structured and simple-to-read format, which can be seamlessly integrated into various applications and systems. Because JSON is a widely used data exchange standard, this functionality streamlines the process of working with the model’s outputs, making it more accessible and practical for developers across different industries and use cases. For more information on generating JSON using the Converse API, refer to Generating JSON with the Amazon Bedrock Converse API.

Additionally, Pixtral Large supports the Converse API and tool usage. By using the Amazon Bedrock API, you can grant the model access to tools that assist in generating responses to the messages you send.

Generate SQL code from a database entity-relationship diagram

An entity-relationship (ER) diagram is a visual representation used in database design to illustrate the relationships between entities and their attributes. It is a crucial tool for conceptual modeling, helping developers and analysts understand and communicate the structure of data within a database.

This example tests the model’s ability to generate PostgreSQL-compatible SQL CREATE TABLE statements for creating entities and their relationships.

We use the following prompt:

prompt = """
You are a database designer. You have created an ER diagram. You need to generate PostgreSQL compatible SQL statements to create the tables"
"""

We input the following ER diagram.

The model response is as follows:

Sure! Below are the SQL statements to create the tables based on the provided 
ER diagram for a PostgreSQL database.

```sql
-- Create the Airport table
CREATE TABLE Airport (
    Airport_ID SERIAL PRIMARY KEY,
    AirportCity VARCHAR(255),
    AirportCountry VARCHAR(255)
);

-- Create the Calendar table
CREATE TABLE Calendar (
    Day_Date DATE PRIMARY KEY,
    Business_Day_YN BOOLEAN
);

-- Create the Flight_Details table
CREATE TABLE Flight_Details (
    Flight_ID SERIAL PRIMARY KEY,
    Source_Airport_ID INT REFERENCES Airport(Airport_ID),
    Destination_Airport_ID INT REFERENCES Airport(Airport_ID),
    Departure_Date_Time TIMESTAMP,
    Arrival_Date_Time TIMESTAMP,
    Airplane_Type VARCHAR(255)
);

-- Create the Travel_Class table
CREATE TABLE Travel_Class (
    Travel_Class_ID SERIAL PRIMARY KEY,
    Travel_Class_Name VARCHAR(255),
    Travel_Class_Capacity INT
);

-- Create the Seat_Details table
CREATE TABLE Seat_Details (
    Seat_ID SERIAL PRIMARY KEY,
    Travel_Class_ID INT REFERENCES Travel_Class(Travel_Class_ID),
    Flight_ID INT REFERENCES Flight_Details(Flight_ID)
);

-- Create the Flight_Cost table
CREATE TABLE Flight_Cost (
    Seat_ID INT REFERENCES Seat_Details(Seat_ID),
    Valid_From_Date DATE,
    Cost DECIMAL(10, 2),
    Valid_To_Date DATE,
    PRIMARY KEY (Seat_ID, Valid_From_Date)
);

-- Create the Flight_Service table
CREATE TABLE Flight_Service (
    Service_ID SERIAL PRIMARY KEY,
    Service_Name VARCHAR(255)
);

-- Create the Service_Offering table
CREATE TABLE Service_Offering (
    Travel_Class_ID INT REFERENCES Travel_Class(Travel_Class_ID),
    Service_ID INT REFERENCES Flight_Service(Service_ID),
    Offered_YN BOOLEAN,
    From_Date DATE,
    To_Date DATE,
    PRIMARY KEY (Travel_Class_ID, Service_ID)
);

-- Create the Passenger table
CREATE TABLE Passenger (
    Passenger_ID SERIAL PRIMARY KEY,
    P_FirstName VARCHAR(255),
    P_LastName VARCHAR(255),
    P_Email VARCHAR(255),
    P_PhoneNumber VARCHAR(255),
    P_Address VARCHAR(255),
    P_City VARCHAR(255),
    P_State VARCHAR(255),
    P_Zipcode VARCHAR(255),
    P_Country VARCHAR(255)
);

-- Create the Reservation table
CREATE TABLE Reservation (
    Reservation_ID SERIAL PRIMARY KEY,
    Passenger_ID INT REFERENCES Passenger(Passenger_ID),
    Seat_ID INT REFERENCES Seat_Details(Seat_ID),
    Date_Of_Reservation DATE
);

-- Create the Payment_Status table
CREATE TABLE Payment_Status (
    Payment_ID SERIAL PRIMARY KEY,
    Payment_Status_YN BOOLEAN,
    Payment_Due_Date DATE,
    Payment_Amount DECIMAL(10, 2),
    Reservation_ID INT REFERENCES Reservation(Reservation_ID)
);
```

These SQL statements will create the tables as per the ER diagram you provided. Make sure to run these statements in the order they are listed to ensure that foreign key constraints are satisfied.

Convert organization hierarchy to structured text

Pixtral Large has the capabilities to understand organization structure and generate structured output. Let’s test it with an organization structure.

We use the following prompt:

prompt = """
Extract organization hierarchy from the given org structure. provide response in a structured json format with below:
- role
- name
- reporting_manager
"""

We input the following organization structure image.

The model response is as follows:

Here is the organization hierarchy extracted from the given org structure in a 
structured JSON format:

```json
[
    {
        "role": "President",
        "name": "John Smith",
        "reporting_manager": null
    },
    {
        "role": "VP Marketing",
        "name": "Susan Jones",
        "reporting_manager": "John Smith"
    },
    {
        "role": "VP Sales",
        "name": "Rachel Parker",
        "reporting_manager": "John Smith"
    },
    {
        "role": "VP Production",
        "name": "Tom Allen",
        "reporting_manager": "John Smith"
    },
    {
        "role": "Manager",
        "name": "Alice Johnson",
        "reporting_manager": "Susan Jones"
    },
    {
        "role": "Manager",
        "name": "Tim Moore",
        "reporting_manager": "Susan Jones"
    },
    {
        "role": "Manager",
        "name": "Michael Gross",
        "reporting_manager": "Rachel Parker"
    },
    {
        "role": "Manager",
        "name": "Kim Dole",
        "reporting_manager": "Rachel Parker"
    },
    {
        "role": "Manager",
        "name": "Kathy Roberts",
        "reporting_manager": "Tom Allen"
    },
    {
        "role": "Manager",
        "name": "Betsy Foster",
        "reporting_manager": "Tom Allen"
    }
]
```

Chart understanding and reasoning

Pixtral Large has the capability to understand and reason over charts and graphs. Let’s test Pixtral Large with a visualization of the portion of motorcycle ownership per country around the world.

We use the following prompt:

prompt = """Examine this visualization:
1. First, describe what this visualization represents:
   - What information is being shown?
   - How is the data displayed?
   - What do the different components represent?
   - What does the size variation indicate?
2. Analyze specific patterns:
   - Which region shows the highest proportion of the first category?
   - Which region shows the highest proportion of the second category?
   - Where do you observe the largest total values?
   - Which areas show the most balanced distribution?
3. Compare regional trends:
   - How do the proportions differ between continents?
   - What patterns emerge between different hemispheres?
   - Are there clear differences between different economic zones?
4. Identify interesting outliers:
   - Which regions stand out from their neighbors?
   - Can you identify any unexpected patterns?
   - Where do you notice significant data variations?
5. Consider geographical and demographic factors:
   - How might local conditions influence these patterns?
   - What socioeconomic factors might explain the variations?
   - Can you identify any correlation between size and proportions?
6. Make comparisons between:
   - Different geographical zones
   - Various population densities
   - Different economic development levels"""

We input the following image.

By Dennis Bratland – Own work, CC BY-SA 3.0, https://commons.wikimedia.org/w/index.php?curid=15186498

The model response is as follows:

Description of the Visualization:
Information Shown: The visualization presents data on the population (in millions) and the proportion of vehicles (cars and motorcycles) across various regions worldwide.
Display Method: The data is displayed using pie charts overlaid on a world map. The size of each pie chart corresponds to the population of the region, with larger pies indicating larger populations.
Components:
Pie Charts: Each pie chart is divided into two segments representing the proportion of cars (light blue) and motorcycles (dark blue).
Size Variation: The size of the pie charts indicates the total population of the region, with larger pies representing higher populations.
Specific Patterns:
Highest Proportion of Cars: Regions like the United States and parts of Europe show a higher proportion of cars.
Highest Proportion of Motorcycles: Southeast Asian countries like Indonesia and Vietnam show a higher proportion of motorcycles.
Largest Total Values: Large pie charts are observed in highly populated regions like India, China, and parts of Southeast Asia.
Most Balanced Distribution: Some regions in Europe and South America show a more balanced distribution of cars and motorcycles.
Regional Trends:
Continental Differences:
North America: Predominantly cars.
Europe: A mix with a slight preference for cars.
Africa: Generally lower total values with varied proportions.
Asia: High population with a significant proportion of motorcycles in Southeast Asia.
Hemispheric Patterns: The Northern Hemisphere has more densely populated regions with varied vehicle proportions, while the Southern Hemisphere shows lower population densities and more balanced distributions.
Economic Zones: Developed regions tend to have higher car proportions, while developing regions, especially in Asia, show higher motorcycle usage.
Interesting Outliers:
Indonesia: Stands out with a very high proportion of motorcycles compared to neighboring regions.
United States: Shows a predominantly high proportion of cars, distinct from other regions.
Unexpected Patterns: Some smaller African countries show unexpectedly high proportions of motorcycles despite lower populations.
Geographical and Demographic Factors:
Local Conditions: Urbanization and infrastructure development influence vehicle preferences.
Socioeconomic Factors: Economic status and affordability play significant roles in vehicle choices.
Correlation: Larger populations in developing regions often correlate with higher motorcycle usage due to affordability and convenience.
Comparisons:
Geographical Zones: Urban areas in developed countries show higher car usage, while rural and developing regions favor motorcycles.
Population Densities: Highly populated areas in Asia show a preference for motorcycles due to traffic congestion and affordability.
Economic Development: Developed economies have higher car proportions, while developing economies rely more on motorcycles.

Conclusion

In this post, we demonstrated how to get started with the Pixtral Large model in Amazon Bedrock. The Pixtral Large multimodal model allows you to tackle a variety of use cases, such as document understanding, logical reasoning, handwriting recognition, image comparison, entity extraction, extracting structured data from scanned images, and caption generation. These capabilities can enhance productivity across numerous enterprise applications, including ecommerce (retail), marketing, financial services, and beyond.

Mistral AI’s Pixtral Large FM is now available in Amazon Bedrock. To get started with Pixtral Large in Amazon Bedrock, visit the Amazon Bedrock console.

Curious to explore further? Take a look at the Mistral-on-AWS repo. For more information on Mistral AI models available on Amazon Bedrock, refer to Mistral AI models now available on Amazon Bedrock.

About the Authors

Deepesh Dhapola is a Senior Solutions Architect at AWS India, specializing in helping financial services and fintech clients optimize and scale their applications on the AWS Cloud. With a strong focus on trending AI technologies, including generative AI, AI agents, and the Model Context Protocol (MCP), Deepesh leverages his expertise in machine learning to design innovative, scalable, and secure solutions. Passionate about the transformative potential of AI, he actively explores cutting-edge advancements to drive efficiency and innovation for AWS customers. Outside of work, Deepesh enjoys spending quality time with his family and experimenting with diverse culinary creations.

Andre Boaventura is a Principal AI/ML Solutions Architect at AWS, specializing in generative AI and scalable machine learning solutions. With over 25 years in the high-tech software industry, he has deep expertise in designing and deploying AI applications using AWS services such as Amazon Bedrock, Amazon SageMaker, and Amazon Q. Andre works closely with global system integrators (GSIs) and customers across industries to architect and implement cutting-edge AI/ML solutions to drive business value.

Preston Tuggle is a Sr. Specialist Solutions Architect with the Third-Party Model Provider team at AWS. He focuses on working with model providers across Amazon Bedrock and Amazon SageMaker, helping them accelerate their go-to-market strategies through technical scaling initiatives and customer engagement

Shane Rai is a Principal GenAI Specialist with the AWS World Wide Specialist Organization (WWSO). He works with customers across industries to solve their most pressing and innovative business needs using AWS’s breadth of cloud-based AI/ML services, including model offerings from top-tier foundation model providers.

Ankit Agarwal is a Senior Technical Product Manager at Amazon Bedrock, where he operates at the intersection of customer needs and foundation model providers. He leads initiatives to onboard cutting-edge models onto Amazon Bedrock Serverless and drives the development of core features that enhance the platform’s capabilities.

Niithiyn Vijeaswaran is a Generative AI Specialist Solutions Architect with the Third-Party Model Science team at AWS. His area of focus is AWS AI accelerators (AWS Neuron). He holds a Bachelor’s in Computer Science and Bioinformatics.

Aris Tsakpinis is a Specialist Solutions Architect for Generative AI focusing on open source models on Amazon Bedrock and the broader generative AI open source ecosystem. Alongside his professional role, he is pursuing a PhD in Machine Learning Engineering at the University of Regensburg, where his research focuses on applied natural language processing in scientific domains.