Pixtral 12B 24.09Open multimodal 12B model handling interleaved images and text with a 128K context window.

4.6 (5)

Reviewed by Daniel Nikulshyn·Updated July 2026

Overview

Pixtral 12B 24.09 is a multimodal model from Mistral AI that processes both images and text within a single sequence, supporting variable image sizes and aspect ratios. It uses a 12-billion-parameter language decoder paired with a vision encoder, enabling tasks like visual question answering, document understanding, chart interpretation, and image captioning. The model accepts up to a 128K token context, allowing multiple images to be interleaved with long-form text in one prompt. Released under an open license, it can be deployed locally or through inference providers, making it suitable for developers building vision-language applications, research workflows, and multimodal agents.

Key features

12B parameter vision-language model
Interleaved image and text inputs
128K token context length
Native variable image size support
Open-weight release
Suitable for OCR, VQA, and captioning

Pricing

Model: Free
Category: LLM
Rating: 4.6 / 5 (5)

Use cases

Multimodal Reasoning

Pixtral 12B is capable of understanding both natural images and documents, achieving state-of-the-art performance on the MMMU reasoning benchmark and surpassing larger models.

Instruction Following

Pixtral 12B excels in instruction following, particularly in multimodal and text-only scenarios, with a 20% relative improvement in text IF-Eval and MT-Bench over the nearest open-source model.

Multimodal Question Answering

Pixtral 12B shows strong abilities in multimodal question answering, including document question answering and chart and figure understanding.

Pros & Cons

Pros

Open weights for self-hosting
Handles multiple images per prompt
Large 128K context window
Flexible image resolutions and aspect ratios

Cons

Requires significant GPU resources
Smaller than frontier closed models
Limited tooling compared to proprietary APIs

Battle record

Across 1 battle in the Pantheon.

1st

2nd

3rd

Last battle

#2
LLM Showdown — February 20, 2026
Feb 20, 2026 · #2 of 3

Reviews

4.6

Average from 5 ratings.

Sanjay Gupta

Jan 7, 2026

Does the job

Pretty happy overall. Open-weight release just works and large 128K context window. but no dealbreakers — I'd recommend it to a friend without hesitating.

Fatima Zahra

Nov 26, 2025

Does the job

Pretty happy overall. Open-weight release just works and handles multiple images per prompt. but no dealbreakers — I'd recommend it to a friend without hesitating.

Naomi Suzuki

Oct 12, 2025

Years in this space

I've evaluated a lot of these over the years. What stands out here is interleaved image and text inputs — handled better than most — and handles multiple images per prompt. Smaller than frontier closed models is my one real gripe. Worth the time if this is your use case.

Tariq Aziz

Oct 7, 2025

Solid for our team

We rolled this out across the team last quarter and open weights for self-hosting. Open-weight release fits neatly into how we already work, and interleaved image and text inputs removed a step we used to do by hand. Smaller than frontier closed models, which is the main caveat, but it has held up under daily use.

Aaliyah Johnson

Aug 30, 2025

Years in this space

I've evaluated a lot of these over the years. What stands out here is 12B parameter vision-language model — handled better than most — and open weights for self-hosting. Smaller than frontier closed models is my one real gripe. Worth the time if this is your use case.

Q&A

How many images and how much text can I include in a single prompt?

Pixtral supports interleaved image and text inputs within a single 128 K token context, allowing any number of images (at their natural resolution) alongside long‑form text in one prompt.

Asked by Lorenzo Bianchi · Nov 20, 2025

Is Pixtral 12B still maintained, and are there newer alternatives?

Pixtral 12B is deprecated and no longer maintained; Mistral AI recommends using its newer, more powerful vision‑language models that supersede Pixtral for production use.

Asked by Carlos Mendoza · Nov 10, 2025

What hardware is needed to run Pixtral 12B effectively?

The model requires substantial GPU memory due to its 12 billion parameters and 400 M‑parameter vision encoder; typical deployments use high‑end GPUs (e.g., A100 40 GB or comparable) to handle the 128 K token context and multiple images.

Asked by Ivo Novotný · Nov 6, 2025

Can I self‑host Pixtral 12B, and under what license?

Yes, Pixtral 12B is released under the Apache 2.0 open‑source license, allowing you to download the weights and run the model locally on your own hardware.

Asked by Priya Nair · Oct 12, 2025