Qwen3-VL 32B OCR API
Model source: Hugging FaceDocument reading for awkward layouts: tables, columns and handwriting.
Qwen3-VL 32B OCR is Alibaba's vision model tuned for documents with difficult layouts: multi-column pages, tables, stamps, handwriting. It returns the text in reading order and the data as structured output. On GPU Flow it costs €0.10 per million input tokens and €0.35 per million output.
Reads in one go
≈ 140 pages
65,536 tokens
Brain size
Mid-sized
32B, NVFP4
Where it runs
Spain (EU)
MIG B200
Understands
Text and images
Vision, OCR
What it is good at
- Tables and multi-column layouts. Keeps rows, columns and reading order intact.
- Handwritten notes and forms. Reads filled-in forms and notes, not only printed text.
- Mixed-language documents. Handles pages that switch language without losing the thread.
- Structured extraction to JSON. Returns the data ready to go into your system, within a context of about 140 pages.
For simple, clean scans at high volume, Nemotron Nano 12B VL is cheaper.
Pricing
€ per million tokens, VAT included
For your technical team
Compatible with the OpenAI SDK: change the base URL and the key, and that is it.
curl https://api.gpuflow.ai/v1/chat/completions \
-H "Authorization: Bearer $GPUFLOW_API_KEY" \
-d '{ "model": "qwen3-vl-32b-ocr",
"messages": [{ "role": "user", "content": "Hola" }] }'Context and limits
- Context
- 65,536 tokens per request, about 140 pages.
- Requests per minute
- 1,000 per key, with bursts up to double; raisable to 5,000 from the dashboard. 5,000 in total per account.
- Tokens per minute
- 2,000,000 input (cached tokens do not count) and 400,000 output, per key.
- Prompt per request
- Up to 256,000 tokens, raisable on request.
- If you go over
- A 429 error with a Retry-After header telling you how many seconds to wait.
Qwen3-VL 32B OCR compared with similar models
| Qwen3-VL 32B OCR | Nemotron Nano 12B VL | |
|---|---|---|
| Best for | Documents with tables, columns and handwriting | Reading scans and images at volume |
| Context | ≈ 140 pages | ≈ 280 pages |
| Input / output, € per million | 0.10 / 0.35 | 0.06 / 0.24 |
| 1,000 chat answers | ~0.16 € | ~0.10 € |
Difficult layouts, tables or handwriting: Qwen3-VL 32B OCR; simple, clean scans at high volume: Nemotron Nano 12B VL.
Frequently asked questions
Is Qwen3-VL 32B OCR free?
The model weights are open source; running them costs GPU time. On GPU Flow you pay €0.10 per million input tokens, and every new account gets €5 of credit when it adds a card.
What does the Qwen3-VL 32B OCR API cost?
€0.10 per million input tokens and €0.35 per million output. VAT included, no monthly fee.
Can I use Qwen3-VL 32B OCR with Claude Code or Cursor?
Yes. Claude Code connects through the Anthropic-compatible endpoint (/v1/messages); Cursor and other tools through the OpenAI-compatible one (/v1/chat/completions).
Where does it run?
On our NVIDIA B200 GPUs, in a data centre in Spain.
Does it read handwriting in Spanish?
Yes, handwriting is one of the layouts it is tuned for, in Spanish and in mixed-language documents. Accuracy depends on how legible the writing and the scan are.
Can it output the table as JSON or CSV?
Yes. Ask for the format in your prompt and it returns the table as JSON or as CSV text.
How are images billed?
Each image counts as input tokens, at the same per-million price as text. There is no separate per-image fee.