Pixtral-12B-2409 - A multimodal AI model by Mistral AI supporting text and image processing with 128k context length.

## Core Architecture of Pixtral-12B-2409 Pixtral-12B-2409 consists of two primary components: - **Decoder**: A 12-billion-parameter transformer model for text processing. - **Visual Encoder**: A 400-million-parameter module for image understanding. The model is designed to handle interleaved image-text data natively. ## Key Capabilities of Pixtral-12B-2409 The model demonstrates strong performance in: - **Multimodal tasks**: Document QA (DocVQA 90.7%), visual question answering (VQAv2 78.6%), and chart analysis. - **Text-only benchmarks**: Maintains competitive performance in pure text generation and comprehension. - **Multilingual support**: Processes 24 languages, including Chinese, English, Japanese, and Korean. ## Performance Comparison with Competing Models - **Vs. GPT-4o Mini**: Pixtral outperforms in specific multimodal benchmarks (e.g., DocVQA, VQAv2) but may lag in text-only tasks like MMLU. - **Vs. Gemma 3**: Direct comparisons are limited due to incomplete benchmark alignment, but Pixtral shows advantages in multimodal reasoning tasks. Note: GPT-4o Mini scores 82% on MMLU (text benchmark), while Pixtral's strengths lie in multimodal applications. ## Hardware Requirements for Deployment The model can run on a **single NVIDIA RTX 4090 GPU (24GB VRAM)**. It is optimized for efficiency and supports: - **Libraries**: `vLLM` (v≥0.6.2) and `mistral-inference` (v≥1.4.1). - **Deployment options**: Local execution or API integration via platforms like La Plateforme. ## Multilingual Support in Pixtral-12B-2409 The model supports **24 languages**, including but not limited to: - Chinese (中文) - English - Japanese (日本語) - Korean (한국어) This broad coverage addresses global use cases. ## Licensing Information Pixtral-12B-2409 is released under the **Apache 2.0 license**, allowing permissive use, modification, and distribution in both research and commercial applications. ## Image Processing Flexibility The model supports: - **Variable image sizes and aspect ratios**. - **Multi-image processing** within its 128k-token context window. - **Image-to-code generation** (e.g., converting diagrams to HTML). ## Benchmark Results Key benchmark scores include: | Task | Metric | Score | |---------------|------------------|-------| | MMMU (CoT) | Accuracy | 52.5% | | Mathvista | Accuracy | 58.0% | | DocVQA | ANLS | 90.7% | | VQAv2 | VQA Match | 78.6% | These highlight its multimodal proficiency. ### Citation sources: - [Pixtral-12B-2409](https://huggingface.co/mistralai/Pixtral-12B-2409) - Official URL Updated: 2025-04-01