Pixtral 12B 24.09

Pixtral-12B-2409 is a 12-billion-parameter multimodal model by Mistral AI, combining a 12B-parameter text decoder with a 400M-parameter vision encoder. It processes interleaved text and images natively, supporting variable image sizes and a 128K-token context window for long-form document analysis or multi-image workflows. The model excels in tasks like chart understanding, OCR, and multilingual reasoning, outperforming similar-sized open models (e.g., Qwen2-VL 7B, LLaVA-OV 7B) and even larger models like Llama-3.2 90B in benchmarks like MMMU (52.5%) and MathVista (58.0%)

• 128K Context Window: Handles long documents or multi-image inputs.
• Variable Image Support: Processes images at native resolution and aspect ratio via a vision encoder.
• Multilingual & Code Capabilities: Supports 80+ coding languages and nuanced multilingual understanding.
• Open Source: Apache 2.0 license for free modification and deployment.
• High Accuracy: Outperforms Claude 3 Haiku and Gemini-1.5 Flash 8B in multimodal benchmarks.
• Vision-to-Code: Generates HTML/CSS from sketches or diagrams

• Image Captioning & OCR: Generate descriptions or extract text from images/documents.
• Data Analysis: Convert charts to Markdown tables or interactive dashboards.
• Document QA: Answer questions from technical manuals or financial reports.
• Academic Research: Summarize papers or analyze scientific diagrams.
• Automation: Integrate with workflows for invoice processing or customer support