What is Qianfan-OCR?
Qianfan-OCR is a document model from Baidu's Qianfan team, released in March 2026 under Apache 2.0 with a technical report on 11 March. It pairs Baidu's Qianfan-ViT vision encoder with a Qwen3-4B language model, 4.74 billion parameters in all, and turns a page image straight into Markdown without a separate layout model. The same weights handle table and formula recognition, key-information extraction, chart and document question answering, and OCR in 192 languages.
Its distinctive feature is Layout-as-Thought. With thinking switched on, the model first writes out bounding boxes, element types from 25 categories and reading order, then the final text, so a single model returns layout data that pipeline systems get from a detector.
It is a separate line from Baidu's PaddleOCR-VL and Unlimited-OCR, which come from other Baidu teams. On Hugging Face it drew 34,433 downloads in the past month and 1,217 likes.
Is Qianfan-OCR worth it?
Qianfan-OCR trades page-parsing accuracy for breadth, and the trade is clear in the independent numbers. OpenDataLab's OmniDocBench v1.6 leaderboard lists it at 93.90, behind GLM-OCR at 95.22 and OvisOCR2 at 96.47, both less than a third of its size. Baidu's headline of 93.12 and first place among end-to-end models is on the older v1.5, where its own table also shows the pipeline-based PaddleOCR-VL 1.5 ahead at 94.50.
Where it leads is the work that follows parsing. Baidu reports a mean of 87.9 across five key-information extraction benchmarks, against 84.2 for Qwen3-VL-235B-A22B and 79.2 for Gemini-3.1-Pro, and 85.2 on the CharXiv reasoning questions where Qwen3-VL-4B scores 48.5. It scores 880 on OCRBench and covers 192 languages, against 109 for PaddleOCR-VL-1.6 and eight for GLM-OCR. All of these are self-reported, though the comparison models are general-purpose and not tuned for documents.
Layout-as-Thought is useful rather than a free gain. Baidu's own figures show thinking mode lowering the OmniDocBench v1.5 overall from 93.12 to 92.64, and its guidance is to enable it only for mixed pages such as exam papers and newspapers. Its value is getting boxes and reading order without running a detector, not better text.
Size is the main cost. At 4.74 billion parameters it reaches 1.024 pages a second on one A100 only with W8A8 quantisation, and 0.503 at full precision, while MinerU 2.5 manages 1.057 in the same table. Baidu's limitations section says the footprint rules out edge devices and CPU-only servers, and lists video, curved text and stylised handwriting as weak spots.
Our take: Qianfan-OCR earns its place when the output feeds a decision, such as reading an invoice total or a chart trend, rather than an archive. Teams that only need clean Markdown get more accuracy per GPU from GLM-OCR or PaddleOCR-VL-1.6.
Where it earns its place
An accounts-payable team at a multinational receiving invoices in a dozen languages
Extract supplier, total and tax fields from invoices without a separate OCR step
Prompt the model for the fields directly from the page image on a vLLM server. Baidu reports 86.5 F1 on Nanonets KIE; the 192-language coverage matters more here than the Markdown score.
An equity research analyst's data team processing broker reports
Read chart values and trends that text-only OCR throws away
Ask chart questions against the page image. Baidu notes that OCR-then-LLM systems score 0.0 on CharXiv because the chart structure is lost; Qianfan-OCR reports 94.0 on the descriptive questions.
A document-AI engineer replacing a layout detector plus recogniser stack
Get bounding boxes, element types and reading order from one model
Enable thinking on mixed pages such as exam papers to get layout output first, and leave it off for single-column documents, where Baidu says results are better and faster without it.
Before you commit
- Setup
- Load baidu/Qianfan-OCR with Transformers' AutoModelForImageTextToText in BF16, or run vllm serve with --trust-remote-code and an hf-overrides flag that maps it to the InternVLChatModel architecture.
- Integrations
- Transformers and vLLM; community GGUF builds for llama.cpp and MLX for Apple silicon; Baidu AI Cloud's Qianfan platform; an agent skill in Baidu's skills repository.
- Data and privacy
- Apache 2.0, with some bundled source files under MIT. Self-hosted, documents stay on your hardware; Qianfan platform retention terms are not stated on the model card.
- Cost at scale
- The BF16 checkpoint is about 9.5 GB. Baidu reports 1.024 pages a second with W8A8 and 0.503 at W16A16 on one A100, roughly 3,700 and 1,800 pages an hour per card.
- Limits
- Visual input up to 4,096 tokens per image in 448-pixel tiles; 32K context by default. Layout-as-Thought is validated only on OmniDocBench v1.5 parsing, not extraction or question answering; too large for edge devices or CPU-only servers.
What’s good
- Mean 87.9 across five key-information extraction benchmarks (self-reported)
- Chart and document question answering in the same model
- Optional layout output with 25 element types, no separate detector
- OCR in 192 languages; Apache 2.0
What’s not
- 93.90 on OmniDocBench v1.6, behind much smaller parsers
- 4.74B parameters; about 1 page a second on an A100 only when quantised
- Thinking mode lowers the overall parsing score
- vLLM needs an architecture override; hosted price not published
Best for
Teams that need field extraction, chart reading and multilingual OCR from one self-hosted model and can spare a data-centre GPU.
Qianfan-OCR alternatives
Qianfan-OCR: common questions
- Is Qianfan-OCR worth it?
- Qianfan-OCR is the model to choose when a page needs reading and understanding in one call; for plain conversion to Markdown, smaller parsers are more accurate. We rate it 4.0 out of 5.
- How much does Qianfan-OCR cost?
- Free under Apache 2.0. Also offered on Baidu AI Cloud's Qianfan platform; the hosted price is not stated on the model card.
- What are the downsides of Qianfan-OCR?
- 93.90 on OmniDocBench v1.6, behind much smaller parsers; 4.74B parameters; about 1 page a second on an A100 only when quantised; thinking mode lowers the overall parsing score; vLLM needs an architecture override; hosted price not published.
- Who is Qianfan-OCR best for?
- Teams that need field extraction, chart reading and multilingual OCR from one self-hosted model and can spare a data-centre GPU.
- How much does Qianfan-OCR cost for a team or in production?
- The BF16 checkpoint is about 9.5 GB. Baidu reports 1.024 pages a second with W8A8 and 0.503 at W16A16 on one A100, roughly 3,700 and 1,800 pages an hour per card.
- Is Qianfan-OCR safe for company data?
- Apache 2.0, with some bundled source files under MIT. Self-hosted, documents stay on your hardware; Qianfan platform retention terms are not stated on the model card.
- What are the best Qianfan-OCR alternatives?
- The closest alternatives we have reviewed are GLM-OCR, PaddleOCR-VL-1.6, Unlimited-OCR 2 and Infinity-Parser2-Flash.
Answers checked Nov 2, 2026.
Make Qianfan-OCR? Show this rating on your site
