Pixtral Large

What it is

Pixtral Large is a multimodal model from Mistral AI, announced on 18 November 2024^[1]. Multimodal means the same model reads both text and images: a user can hand it a scanned page, a chart or a photograph together with a question, and the answer is generated from both. It was built on Mistral Large 2, the company's text model, rather than designed from scratch as a vision system^[1].

The model's name is often shortened to Pixtral in casual use, and its API identifier is pixtral-large-2411^[3].

How it is built

The model has two parts. A multimodal decoder with 123 billion parameters does the language work, and a vision encoder with 1 billion parameters turns images into representations the decoder can read, for roughly 124 billion parameters in total^[1]. In this arrangement the encoder is a front end: it converts a picture into a sequence of internal vectors, and the decoder treats those vectors as if they were extra words in the prompt, generating the answer one piece at a time.

The context window is 128K tokens, which Mistral describes as enough to hold at least 30 high-resolution images^[1]. A context window is the amount of input a model can consider at once; a large one lets a user submit a whole multi-page document, with its figures, in a single request instead of feeding it in page by page.

Reported results

On MathVista, a benchmark of mathematical reasoning that requires reading figures and charts, Pixtral Large scored 69.4%^[1]. Mistral also stated that the model surpassed GPT-4o and Gemini 1.5 Pro on ChartQA and DocVQA^[1], two benchmarks that respectively test question answering over charts and over scanned or photographed documents. These comparisons are the developer's own claims rather than results from an independent evaluation, and should be read as such.

How it was made available

Mistral released the weights on Hugging Face under the Mistral Research License, and offered a commercial licence for production use^[1]. Open weights mean the parameters can be downloaded and run by the user, in contrast to a service reachable only through an API; the licence terms, not the file format, determine what the download may be used for.

Mistral's changelog lists mistral-large-2411 as released on the same day as Pixtral Large^[4], so a text-model update and the multimodal version shipped together.

The model was also put to work in the company's own products. In Le Chat's November 2024 update, document and image understanding were powered by Pixtral Large^[2] — the assistant could read an uploaded file or picture and answer questions about it.

Retirement

Mistral's documentation records that the model was deprecated on 27 February 2026 and retired on 31 May 2026, replaced by Mistral Medium 3.5^[3]. The Pixtral Large page states the model is deprecated and has been replaced by newer vision and multimodal models^[1].

Deprecation is a notice period: the model still answers requests but is no longer the one to build new work on. Retirement is the end of that period, after which the identifier stops serving.

Where it sits

Pixtral Large belongs to the generation of multimodal models that shipped as downloadable weights rather than as closed services only, which matters to teams that need to run inference on their own hardware or keep data in house. It also shows how quickly the line moved: from announcement in November 2024 to retirement in May 2026 is roughly a year and a half, with a successor named and a changelog entry marking the day it arrived alongside a text-model release^[1][3][4].