Qianfan-OCR
Open Weights
Baidu
Qianfan-OCR is a 4B-parameter end-to-end document-intelligence model from Baidu's Qianfan team, published under Apache 2.0. It unifies document parsing, layout analysis and document understanding in one vision-language architecture: a Qianfan-ViT encoder (24 layers, AnyResolution up to 4K, 256 visual tokens per 448x448 tile) feeds a Qwen3-4B language model through a two-layer MLP adapter, converting images straight to Markdown. An optional Layout-as-Thought phase, triggered by think tokens, first emits bounding boxes, element types and reading order before the final output. Supported tasks span parsing, tables, formulas, charts, document QA, handwriting and key-information extraction across 192 languages.
Strengths
- 93.12 overall on OmniDocBench v1.5, the leading end-to-end result in the vendor's table, ahead of DeepSeek-OCR-v2 at 91.09 and Gemini-3 Pro at 90.33
- 79.8 on olmOCR Bench and 880 on OCRBench, the latter ahead of Qwen3-VL-4B's 873
- Key-information-extraction mean of 87.9 across five public benchmarks, ahead of Qwen3-VL-235B-A22B (84.2) and Gemini-3.1-Pro (79.2)
- Optional Layout-as-Thought mode returns explicit bounding boxes, element types and reading order alongside the parse, recovering layout analysis inside an end-to-end model
- 192 languages spanning Latin, Cyrillic, Arabic, South/Southeast Asian and CJK scripts
- Strong chart reasoning: CharXiv_DQ 94.0, CharXiv_RQ 85.2, ChartQA 88.1 and ChartBench 85.9, where two-stage OCR-plus-LLM systems score 0.0 on CharXiv
- 1.024 pages/second with W8A8 quantisation on a single A100 under vLLM 0.10.2
Caveats
- On the vendor's own OmniDocBench v1.5 table the PaddleOCR-VL 1.5 pipeline scores higher (94.50 vs 93.12), so the #1 claim is scoped to end-to-end models only
- It loses to the much smaller Qwen3-VL-4B on several general rows: DocVQA 92.8 vs 94.9, TextVQA 80.0 vs 81.8, and OCRBenchv2 English 56.0 vs 60.68
- No release date appears anywhere on the model card as text - only a 2026 bibtex year - so none is recorded here
- Layout-as-Thought is not free: the vendor advises disabling it on homogeneous documents such as single-column text and simple forms, where it gives worse results and higher latency
- vLLM serving requires --trust-remote-code plus an hf-overrides architecture swap to InternVLChatModel, and unquantised W16A16 throughput drops to 0.503 pages/second
- Vendor
- Baidu
- License
- Apache-2.0
- Modalities
- textimage
| Benchmark | Score | Max |
|---|---|---|
| OmniDocBench v1.5 | 93.1 | 100.0 |
| olmOCR Bench | 79.8 | 100.0 |
| OCRBench | 880.0 | 1000.0 |
| Key Information Extraction (overall) | 87.9 | 100.0 |
| DocVQA | 92.8 | 100.0 |
Capabilities
Vision
Audio
Video
Tool Use