Outperforming the crowd: Google Gemini 2.5 Pro becomes the first AI model to fully understand the layout of PDF and can accurately reference it. The latest report points out that Google's Gemini 2.5 Pro model can accurately analyze the visual structure of PDF documents and achieve accurate visual reference functions, becoming the first AI model that can fully understand the layout of PDF. IT Home Note: Google released the Gemini 2.5 Pro experimental model to paying users and developers on March 25. Just 4 days later, Google opened it to users around the world through free web applications. Gemini 2.5 Pro not only extracts textual content from PDF documents, but also understands its visual layout, including charts, tables, and overall layout. Google stated in the developer documentation that this model has Native Vision capabilities, supports processing of up to 3,000 PDF files (each file is limited to 1,000 pages or 50MB), and has a large context window of 1 million tokens, with plans to expand to 2 million tokens in the future. Sergey Filimonov, co-founder of AI startup Matrisk, especially praised Gemini 2.5 Pro for its performance on PDF visual references. Filimonov pointed out that traditional text segmentation methods will cut off the user's visual connection with the original text, making it impossible to visually verify the source of the information. Even in ChatGPT, clicking on a quote only downloads a PDF, forcing users to judge for themselves whether the model is illusory, which seriously damages user trust. In the past, citing document content often only highlighted large sections of irrelevant text, with extremely low accuracy. Gemini 2.5 changes that, not only mapping extracted text fragments back to the exact location of the original PDF, but also targeting specific sentences, table cells, and even images with unprecedented precision. This technological breakthrough provides users with intuitive visual feedback. For example, when asking about housing rate changes, the system can directly highlight relevant data in the document (such as a 15.4% rate change) and mark the source. This level of clarity and interactivity is unmatched by existing tools. Gemini 2.5 not only optimizes the existing process, but also opens up a new document interaction mode. In comparison, Gemini 2.5 is significantly ahead of other models, such as OpenAI s GPT-4o (0.223) and Claude 3.7 Sonnet (0.210), with an IoU (intersection over union) accuracy of 0.804, demonstrating amazing spatial understanding capabilities. Provider Model IOU Summary Gemini 2.5 Pro 0.804 Very good Gemini 2.5 Flash 0.614 Sometimes good Gemini 2.0 Flash 0.395 OpenAI gpt-4o 0.223 OpenAI gpt-4.1 0.268 OpenAI gpt-4.1-mini 0.253 Claude 3.7 Sonnet 0.210 Gemini 2.5 The potential goes far beyond text positioning. It also extracts structured data from PDFs while clearly labeling the source of each data, solving the trust barrier that arises from unclear data origins in downstream decision-making