Model comparison ยท 2026-09-22
Multi-page document analysis with DeepSeek V4 Flash Vision
How to handle page ordering, image submission format, and the 1M-token context window when processing a document dozens of pages long instead of a single image.
A multi-page document is a different problem than a single image
The core flow described in the DeepSeek V4 Flash Vision guide relies on attaching a single image to a request and asking a question. When processing a multi-page document (say, a 20-page report or a scanned contract), two extra decisions come into play: what order to give the pages to the model in, and how to track which page a piece of model output belongs to.
The 1M-token context window is an advantage here: an entire multi-page document can usually be sent in a single request without exceeding the context limit. But fitting in context does not mean the model correctly connects relationships across pages; that needs separate verification.
Make page order and source explicit in the request
Labeling which page each image belongs to in text as well (for example, with a 'Page 3' tag) makes it easier for the model to cite a page reference in its output. Without that labeling, even if the model finds the right information, it can be inconsistent about naming which page it came from, which makes verifying the finding against the document harder.
If a scanned form has handwriting or a low-resolution section, resending just that page at higher resolution or cropped in a separate request is a faster way to verify than reprocessing the whole document.
- Label every image with its page number; this makes citing a reference in output easier for the model.
- 1M context carries most documents in one request, but accuracy still needs separate verification.
- Re-verify one low-quality page with a separate, higher-resolution request rather than reprocessing everything.
Cost is calculated differently on multi-page requests
An image maps to proportionally more tokens than text; input token count grows quickly when a 20-page document is sent in a single request. If a peak-hour multiplier applies (as covered in the DeepSeek V4 Flash Vision guide), scheduling multi-page batch jobs for off-peak hours can noticeably reduce cost.
Frequently asked questions
How many page images can I send in a single request?
There is no fixed page limit; the actual limit is that total image plus text tokens must not exceed the 1M-token context window. Token consumption rises with page count, so consider splitting very large documents.
Does the model always cite the correct page?
If you send images without page-number labels, the model can mix up references. Labeling each image explicitly increases the chance of getting a consistent page reference in the output.