Pixtral 12B
A 12B multimodal model that reads images and text without losing text skills
Pixtral 12B is a 12-billion-parameter multimodal language model from Mistral AI, released in October 2024 under Apache 2.0. It pairs a new vision encoder with a decoder based on Mistral Nemo 12B, handling many images in a 128K-token context while keeping strong text performance.
- 12B multimodal decoder based on Mistral Nemo 12B
- New 400M vision encoder trained from scratch
- Handles any number of images in a 128K-token context
- 52.5% on MMMU, per Mistral AI's own evaluation
- Released under Apache 2.0; now deprecated by Mistral AI
What it is
Pixtral 12B is a multimodal language model built by Mistral AI. It takes in text and images and produces text. The model has 12 billion parameters in total: a 400M-parameter vision encoder trained from scratch and a 12B multimodal decoder derived from Mistral Nemo 12B. Mistral AI designed it to handle natural photographs and documents, including charts, figures and question answering over documents. It accepts images at their native resolution and aspect ratio, and it can process any number of images inside a 128K-token context window. The company positions it as a drop-in replacement for Mistral Nemo 12B, keeping text abilities such as coding and maths while adding vision.
How it came about
Mistral AI announced Pixtral 12B in 2024, with a technical paper posted to arXiv on 9 October 2024 and revised the following day. The paper lists 42 authors, among them Guillaume Lample, Albert Q. Jiang, Baptiste Rozière and Arthur Mensch. The model was made available on La Plateforme and Le Chat. Mistral AI has since marked the model as deprecated, saying it is no longer maintained and has been replaced by newer vision and multimodal models.
Inside the model
Pixtral has two parts. A vision encoder turns each image into tokens, one per 16x16 patch, at the image's own resolution and ratio. Those tokens are flattened into a sequence, with special [IMG BREAK] and [IMG END] markers between rows and at the end, so images of different shapes stay distinguishable even when they use the same token count. A multimodal transformer decoder then predicts the next text token from a mix of text and images. Training used interleaved image and text data. The decoder is based on Mistral Nemo 12B, and the whole system works within a 128K-token window.
What it can do
Mistral AI reports 52.5% on the MMMU reasoning benchmark, which it says beats several larger models. In its own re-evaluation with a shared harness and prompt, Pixtral scored 58.0 on Mathvista, 81.8 on ChartQA, 90.7 on DocVQA and 78.6 on VQAv2. The company says it outperforms open models of similar size, including Llama-3.2 11B and Qwen2-VL 7B, and beats Llama-3.2 90B while being seven times smaller. It also reports a 20% relative gain over the nearest open model on text IF-Eval and MT-Bench. The team created MM-IF-Eval and MM-MT-Bench to test multimodal instruction following.
How to use it
The weights were released under the Apache 2.0 licence. Mistral AI offered the model through La Plateforme and its Le Chat assistant. The company says the prompts used for its benchmarks would be open-sourced, along with MM-MT-Bench. Pixtral 12B is now flagged as deprecated and no longer maintained.
Reception and debate
Mistral AI's own comparisons place Pixtral above open models of similar scale and, in some cases, above closed ones such as Claude 3 Haiku, while matching or beating much larger models like LLaVA OneVision 72B on multimodal benchmarks. The company notes that its scores come from a single evaluation harness with one prompt per dataset, chosen to reproduce published results for GPT-4o and Claude 3.5 Sonnet. No independent reception is described here.
Why it matters
Pixtral 12B was presented as a model that adds vision without giving up text quality, unlike many open multimodal systems of its time. The team also released MM-MT-Bench, a benchmark for vision-language models in practical use, plus code and analysis for standardised multimodal evaluation. Mistral AI now points users to its newer vision and multimodal models instead.
Questions people ask about Pixtral 12B
What is Pixtral 12B?
It is a 12-billion-parameter multimodal language model from Mistral AI that accepts text and images and outputs text. It was released in October 2024 and is aimed at tasks like chart reading, document question answering and multimodal reasoning.
How does it handle images?
A vision encoder trained from scratch converts each image into tokens at its natural resolution and aspect ratio, one token per 16x16 patch. Special markers separate rows and end the image, so different shapes stay distinct. Any number of images fit in the 128K-token window.
Is it open?
Yes. Mistral AI released the weights under the Apache 2.0 licence, and offered the model on La Plateforme and Le Chat. The company has since deprecated it and says it is no longer maintained.
How does it compare to other models?
In Mistral AI's own benchmark runs, Pixtral 12B outperforms open models near its size, such as Llama-3.2 11B and Qwen2-VL 7B, and beats Llama-3.2 90B despite being seven times smaller. It reports 52.5% on MMMU.
Good reading and sources
There is no English Wikipedia article on Pixtral 12B yet, so the facts on this page come from the sources listed above: mistral.ai/news/pixtral-12b, arxiv.org/abs/2410.07073. Some are the subject's own pages, so its claims about itself are reported as its claims. EveryLLM wrote the text in its own words, with AI assistance, and compares it with the sources to make sure no passage is copied.
It may contain mistakes, and figures change over time. Nothing here is investment or buying advice. If something is wrong, tell us. Last refreshed 2026-11-14.