Pixtral 12B

What it is

Pixtral 12B is the first multimodal model from Mistral AI, announced on 17 September 2024^[1]. Multimodal here means the model takes images as well as text as input and answers in text: hand it a chart, a scanned page or a photograph, add a question, and it replies in words. The model is a 12-billion-parameter multimodal decoder built on Mistral NeMo, paired with a 400-million-parameter vision encoder trained from scratch^[1].

Two parts, two jobs

The split is the usual one for this kind of system. The vision encoder looks at the image and turns it into a sequence of internal representations that the language side can read; the decoder, which holds most of the parameters and most of the language ability, reads those representations together with the text of the prompt and produces the answer. Training the encoder from scratch rather than borrowing a pretrained one means the image side was built specifically to fit this decoder^[1].

Native resolution

The detail most often pointed at is how the model handles image size. Pixtral processes images at their native resolution and aspect ratio, turning the picture into one token per 16×16 pixel patch^[1]. Many vision-language models first squash an image to a fixed square, which distorts a wide table or a tall scanned page; keeping the original shape avoids that, at the cost of a token budget that grows with the area of the image. A 512×512 picture, for instance, comes to 32×32 patches, or 1,024 tokens. The context window is 128K tokens and can hold several images at once^[1]. Mistral's model card describes the same capability as support for variable image sizes, and for multiple images per message or across a multi-turn conversation^[3].

Benchmarks

Mistral published five scores for the model: MMMU with chain-of-thought prompting at 52.5%, MathVista 58.0, ChartQA 81.8, DocVQA 90.7 and VQAv2 78.6^[1]. These benchmarks test different things. MMMU poses college-level questions across many disciplines that come with images; MathVista tests mathematical reasoning where the problem is presented visually; ChartQA asks questions about charts; DocVQA asks questions about scanned documents; VQAv2 is general visual question answering about everyday photographs. Reading them together gives a rough picture: strong on documents and charts, mid-range on general image understanding and on hard multi-discipline reasoning.

The model card's comparison table lists Pixtral 12B with the highest MMMU, MathVista, ChartQA and VQAv2 scores among the models it compares^[3]. It is not ahead everywhere: the same table lists Qwen2 7B VL at 94.5 on DocVQA, above Pixtral 12B's 90.7^[3]. Comparisons of this kind depend heavily on which models are placed in the table, and the figures above are the ones published by the maker.

Release

The weights were released under the Apache 2.0 licence and made available on Hugging Face, in Le Chat and on La Plateforme under the name pixtral-12b-2409^[1]. Apache 2.0 is a permissive open-source licence: it lets others use, modify and redistribute the weights, including commercially, as long as they keep the licence notices. Putting a multimodal model out under such terms, rather than behind an API only, meant outside developers could download it, run it themselves and fine-tune it.

Retirement

The hosted model had a short working life. It was deprecated on 2 December 2025 and retired on 31 December 2025, replaced by Ministral 3 14B^[2]. Deprecation is the warning stage, in which a model still answers but users are told to migrate; retirement is the point at which requests to the old name stop being served.

Where it sits

As Mistral's first multimodal model, Pixtral 12B marks the point at which the company's model line went from text to images. The combination it introduced — open weights, native-resolution image handling, a long context that can hold several images — became the template for the multimodal releases that followed. Its retirement also illustrates how quickly the serving side of a released model moves on: the weights remain downloadable under their licence, but the hosted endpoint is pointed at a successor within a bit more than a year.