Pixtral Large (24.11)

// Overview

Pixtral Large (24.11) is Mistral AI's frontier-class multimodal model, released on 18 November 2024. It is the second model in the Pixtral line (after Pixtral 12B) and pairs a 1-billion-parameter vision encoder with the 123-billion-parameter Mistral Large 2 decoder, for roughly 124B parameters in total. Mistral positioned it as a model that adds frontier-level image, document, and chart understanding on top of Mistral Large 2's text capabilities.

The model reads both text and images and ships with a 128K-token context window — enough, per Mistral, to fit a minimum of 30 high-resolution images alongside text in a single request. That makes it suited to multi-page document understanding, chart and diagram reading, and OCR-style tasks, not just single-image captioning. It is distributed with open weights: the instruction-tuned checkpoint (mistralai/Pixtral-Large-Instruct-2411) is downloadable from Hugging Face under the Mistral Research License, with a separate Mistral Commercial License for production use.

Pixtral Large was served via Mistral's API as pixtral-large-2411 (alias pixtral-large-latest) and in le Chat. Mistral has since deprecated it: its documentation lists a deprecation date of 27 February 2026 and points new projects to Mistral Medium 3.5 as the recommended replacement. The open weights remain available for research and self-hosting (Mistral recommends vLLM for inference).

// Benchmarks

1. MathVista (CoT)69.4
2. MMMU (CoT)64
3. ChartQA (CoT)88.1
4. DocVQA (ANLS)93.3
5. VQAv2 (VQA Match)80.9
6. AI2D (BBox)93.8
7. MM MT-Bench7.4

Scores on a 0–100 scale (25-point gridlines); higher is better. Each benchmark links to its published source.

// Strengths

• Strong document, chart, and diagram understanding: 93.3 on DocVQA (ANLS) and 88.1 on ChartQA (CoT) per Mistral's model card
• Leading visual mathematical reasoning for its generation — 69.4 on MathVista (CoT)
• Open weights under the Mistral Research License, so it can be self-hosted and fine-tuned (vLLM-supported)
• Large 128K context that accommodates roughly 30 high-resolution images plus text in one request
• Built on Mistral Large 2, so it retains that model's text-only quality while adding vision

// Best for

• Multi-page document and PDF understanding, including OCR-style extraction
• Reading and reasoning over charts, tables, and diagrams
• Visual question answering over natural images and screenshots
• Math and STEM problems presented as images (figures, plotted data)
• Self-hosted multimodal deployments where downloadable open weights are required

// How to access

// Pixtral — every version

The full lineage of the Pixtral line, newest first. Every version has its own page — click any to compare specs, benchmarks and pricing.