Pixtral: Mistral brings vision to European open models
In one sentence Mistral releases Pixtral 12B (September, Apache 2.0) and Pixtral Large 124B (November): first competitive European multimodal models. Strong focus on document understanding and OCR.
Mistral, the French startup that makes competitive open models, enters multimodal. First it releases Pixtral 12B in September 2024 (Apache 2.0 license, truly free), then in November it ships Pixtral Large at 124 billion parameters.
What it does: you show it a photo, a scanned document, a screenshot, a chart, and ask questions. It recognizes text (OCR), understands technical diagrams, reads receipts, describes scenes.
Key point: it's one of the first open vision models you can download and use in the EU without political license constraints, and it's competitive with Llama 3.2 vision and Claude 3 Haiku on document benchmarks. For European enterprise use cases (banks, public sector, legal) that cannot send data to OpenAI or Anthropic, it becomes a concrete option.
Companies
Mistral AI
Tools
Pixtral 12B, Pixtral Large
Tags
Sources