Qwen2.5 VL 72B Instruct: Documents, Charts and Grounding

The common mistake with vision models is testing them on photographs. Ask one to describe a dog in a park and almost any modern model will do fine. The jobs that pay for themselves are duller: read this invoice, extract this table, find the submit button on this screenshot, tell me what this chart says about Q3.

Qwen2.5 VL 72B Instruct was built with those jobs in mind. Qwen released the Qwen2.5-VL family on 26 January 2025 in 3B, 7B and 72B sizes, and the launch material leans heavily on documents, charts, layout and screen interaction rather than general scene description. This guide covers the 72B instruct model.

What it is designed to do

Qwen's release post lists five headline capabilities:

• Document parsing — multilingual, multi-orientation OCR, with output in a custom QwenVL HTML format that keeps layout information.
• Grounding — locating objects with bounding boxes or points and returning them as JSON, using the image's real pixel coordinates rather than normalised values.
• Charts and graphics — reading text, icons, charts and layouts inside images.
• Long video — understanding videos over an hour long and localising events to the second.
• Visual agent work — operating phone and desktop interfaces without task-specific fine-tuning.

Architecturally, the vision encoder is a dynamic-resolution ViT trained from scratch, with window attention in all but four layers to keep cost down on large images. The language side is a 72B dense model; the weights come to about 73B parameters in total.

The benchmark table, read with care

The model card reports these figures for the 72B model:

• DocVQA (val) — 96.4
• ChartQA (test) — 89.5
• OCRBench — 885
• CC-OCR — 79.8
• MathVista (mini) — 74.8
• MMMU (val) — 70.2
• Video-MME — 73.3 without subtitles, 79.1 with
• ScreenSpot — 87.1, AndroidWorld — 35% success

The pattern is clear. Document and chart reading scores are very high; MMMU, which tests college-level reasoning over images, is respectable but lower; and the agentic AndroidWorld success rate shows that driving a phone end to end was still mostly unsolved. Locating an element on a screen (ScreenSpot) is much easier than completing a multi-step task with it.

These are vendor-run numbers. DocVQA in particular is close to saturated, so the difference between 94 and 96 tells you little about your own scanned forms. Best model for vision tasks covers which benchmarks transfer to which jobs.

Resolution is your cost dial

Because the encoder works at native resolution, an image becomes a variable number of tokens. The model card gives a default range of 4 to 16,384 visual tokens per image, configured through min_pixels and max_pixels, with dimensions rounded to multiples of 28. Its example setting caps an image at 1,280 tokens.

This is the single most important setting to get right. A full-page scan at high resolution can cost thousands of input tokens before you have written a word of prompt. Too low and small print becomes unreadable. Start with the example cap, then raise it only for document types where you can measure an accuracy gain.

OpenRouter lists the model at $0.80 per million input tokens and $1.00 per million output at the time of writing, with a 128,000-token context. Image tokens bill as input, so input price matters far more here than for a chat model. The model card itself ships a 32,768-token configuration and warns that YaRN scaling significantly hurts temporal and spatial localisation, so if you self-host, extend the window carefully.

The licence is not Apache 2.0

The 72B model is tagged with the Qwen licence on Hugging Face, not Apache 2.0. The agreement grants a royalty-free commercial licence, with three conditions worth knowing:

• If your product or service exceeds 100 million monthly active users, you must request a separate licence from Alibaba Cloud.
• Distributions must carry the licence and a notice file crediting Qwen.
• If you use the model or its outputs to train or improve a model that is distributed or made available, you must display "Built with Qwen" or "Improved using Qwen".

For most teams none of this blocks anything. The last clause matters if you plan to distil document-extraction data from it into your own model. The model licence comparison sets these terms beside the other common licences.

Where it fits now

The obvious successor is the Qwen3-VL line, which includes mixture-of-experts sizes as well as dense ones. If you are starting fresh, the Qwen3 VL 32B guide is the nearest like-for-like comparison: a dense model at under half the size, so worth testing first on cost alone.

Qwen2.5 VL 72B remains a sound choice where you already have extraction prompts tuned on it, where the QwenVL HTML output format is wired into your pipeline, or where you need a model whose behaviour on your documents you have already measured.

How to evaluate it on your documents

1. Collect fifty real documents covering your worst cases: skewed scans, dense tables, handwriting, mixed languages.
2. Hand-label the fields you need, or the table cells, once.
3. Run the model at two resolution caps and record field-level accuracy and input tokens per document.
4. Check every failure. Misread digits and dropped table rows are different problems with different fixes.

Cost per correctly extracted document is the number that decides it, not DocVQA.