Pixtral Large

Processes up to 30 high-resolution images interleaved with text in a single 128K-token context window using a dedicated 1B-parameter vision encoder that tokenizes visual input at ~4.3K tokens per image average. The vision encoder feeds into a 123B multimodal decoder backbone (Mistral Large 2) that performs joint reasoning over image and text tokens, enabling sequential image-text conversations where images can appear anywhere in the conversation flow rather than only at the beginning.

Extracts and reasons over text content from scanned documents, receipts, invoices, and forms using integrated optical character recognition (OCR) combined with visual reasoning. The model processes document images through the vision encoder to identify text regions, extract character sequences, and understand document structure (tables, sections, headers), then answers natural language questions about extracted content. Demonstrated on multilingual documents (Swiss German/French receipts) indicating cross-language OCR capability.

Processes documents and images containing text in multiple languages, with demonstrated support for Swiss German and French. Vision encoder extracts text regardless of language, and language decoder applies multilingual understanding to answer questions and extract information. Specific language support list not documented, but multilingual OCR capability confirmed through receipt processing examples.

Analyzes charts, graphs, and data visualizations to extract numerical values, identify trends, and perform mathematical reasoning over visual data. The model processes chart images through the vision encoder to recognize chart types (bar, line, scatter, pie, etc.), extract axis labels and data points, then applies mathematical reasoning to answer questions like 'what is the trend?' or 'calculate the average'. Demonstrated on ChartQA and MathVista benchmarks with claimed superiority over GPT-4o and Gemini-1.5 Pro.

Performs multi-step visual reasoning over natural images containing objects, scenes, spatial relationships, and contextual information. The vision encoder tokenizes image content into visual tokens that the 123B language decoder processes using attention mechanisms to identify objects, understand spatial layouts, reason about relationships, and answer complex questions requiring scene understanding. Supports reasoning chains that decompose visual understanding into steps.

Enables the model to invoke external tools and functions based on visual understanding, allowing image analysis to trigger downstream actions or API calls. The model can analyze an image, extract relevant information, and call functions with extracted parameters (e.g., 'analyze receipt image → extract vendor name, amount, date → call accounting API with structured data'). Implementation details of tool schema binding and function registry not documented.

Maintains full text-only capabilities of Mistral Large 2 base model including code generation, reasoning, summarization, and general language tasks. The 123B language decoder processes text tokens independently of vision encoder, enabling pure text interactions and leveraging Mistral Large 2's instruction-tuning for diverse language tasks. 128K context window applies to text-only conversations as well.

Available as open-weights model under Mistral Research License (MRL) and Mistral Commercial License, enabling self-hosted deployment on private infrastructure without API dependency. Model distributed in unspecified format (likely safetensors or GGUF) for download and local inference. Supports both research/educational use (MRL) and commercial deployment (Commercial License), though specific license terms and restrictions not detailed in documentation.

Available through Mistral API as `pixtral-large-latest` endpoint supporting standard API patterns including streaming responses, batch processing, and integration with Mistral's ecosystem tools. API endpoint abstracts hardware deployment complexity and provides managed inference with automatic scaling. Pricing model not documented in provided materials (pricing page references Le Chat subscription but not per-token API costs).

Accessible through Mistral's Le Chat web interface, providing browser-based access to Pixtral Large without API integration or local deployment. Le Chat handles image upload, conversation management, and response rendering. Subscription-based access model with tiers referenced in pricing documentation, though specific tier features and costs not detailed for Pixtral Large specifically.

Supports 128K token context window enabling extended conversations with multiple images and long text passages. Context window is shared between image tokens (approximately 4.3K tokens per high-resolution image) and text tokens, allowing up to 30 high-resolution images or proportionally more text. Enables multi-turn conversations where previous context is maintained across turns without re-uploading images.