What it is
PaddleOCR-VL-1.6 is a compact open document-parsing model from Baidu's PaddlePaddle team, released on 28 May 2026 under Apache 2.0, with a technical report on 2 June 2026. It reads text, tables, formulas, charts and seals from page images in 109 languages, and is a drop-in replacement for PaddleOCR-VL-1.5. The model pairs a NaViT-style visual encoder with Baidu's ERNIE-4.5-0.3B language model, about 0.96 billion parameters in total.
The upgrade targets regions where the previous version made mistakes, with extra training data for those areas and reinforcement learning after training. The team reports gains on tables, ancient Chinese documents, rare characters, seals and charts. Whole-page parsing runs through the PaddleOCR pipeline, which adds a separate layout-analysis model in front of the VLM.
Our review
PaddleOCR-VL-1.6 offers the best accuracy per parameter of any open document parser, with the caveat that the accuracy figures are its makers'. The team reports 96.33% on OmniDocBench v1.6 and new highs on v1.5 and its own Real5-OmniDocBench, which tests scans, warped pages, screen photos, poor lighting and skew. An earlier PaddleOCR-VL already beat DeepSeek-OCR 2 on v1.5 in DeepSeek's own paper, 92.86 to 91.09, which lends the lineage some outside support.
The scoring method deserves a second look. The card says rival numbers come from the public OmniDocBench leaderboard except for Gemini 3 Pro, Qwen3-VL-235B and PaddleOCR-VL-1.6 itself, which the team ran independently. A buyer comparing parsers should rerun a sample of their own documents rather than rely on a table assembled that way.
Deployment is where the cost lives. Full-page results need PaddlePaddle 3.2.1 or later plus the PaddleOCR 3.6 pipeline; the Transformers route recognises single elements but not whole pages. The docs say vLLM needs an NVIDIA card of compute capability 8.0 or higher (RTX 30 series, A10, A100) and warn that T4 and V100 cards are prone to timeouts and out-of-memory errors, and the official Docker images, about 8 GB, are pulled from a Baidu registry in Beijing, which some security teams will want to mirror first.
Size remains the strongest argument. At about 0.96 billion parameters, roughly 1.9 GB in BF16, it is a third the size of the 3-billion-parameter decoders in DeepSeek-OCR 2 and Unlimited-OCR, so one mid-range GPU can serve high page volumes. Unlimited-OCR still handles 40-page documents in a single pass, which PaddleOCR-VL's page-by-page pipeline does not attempt.
Where it earns its place
A data engineer at an insurer processing scanned claim forms
Turn photographed and skewed forms into structured tables
Run the PaddleOCR pipeline with the VLM behind a vLLM server on one A10 or RTX-class card. Weights are free; the job is keeping PaddlePaddle and vLLM versions aligned, and testing on your own forms since every score is the team's.
An ML engineer building a RAG index over research PDFs
Extract formulas and tables accurately enough to search them
Page-level parsing outputs formulas and tables as structured text for chunking. Each page is processed separately, so long tables that span pages need stitching afterwards.
An archive or library digitising historical Chinese documents
Recognise rare characters and seals in old manuscripts
Version 1.6 adds targeted training for ancient documents, rare characters and seals, and the Apache 2.0 licence permits fine-tuning on a private collection without fees.
Before you commit
- Setup
- Install paddlepaddle-gpu 3.2.1+ and paddleocr[doc-parser] 3.6+, or run Baidu's Docker image (about 8 GB; 10 GB offline build); macOS needs Docker.
- Integrations
- PaddleOCR CLI and Python API, VLM serving through vLLM, SGLang or FastDeploy, Transformers 5 for element-level tasks, GGUF builds, and a self-hosted API service via Docker Compose.
- Data and privacy
- Self-hosted only, so documents stay on your hardware. Apache 2.0 permits commercial use and fine-tuning with no revenue cap.
- Cost at scale
- Weights are free and about 1.9 GB in BF16, so a single RTX 30-series or A10 card serves it; no first-party pages-per-second figure or hosted price is published.
- Limits
- Page-by-page processing; Transformers support covers single elements only; vLLM on T4 or V100 is prone to timeouts and out-of-memory errors; directories of PDFs cannot be passed in one call.
What's good
- 96.33% on OmniDocBench v1.6, per its makers
- About 0.96B parameters, roughly 1.9 GB in BF16
- Strong on tables, formulas, seals and real-world photos
- Apache 2.0, with official GGUF builds
What's not
- Headline scores are self-reported
- Full-page parsing needs the PaddlePaddle stack
- vLLM serving needs a compute-capability 8.0+ GPU
- No first-party hosted API or published throughput
Best for
Teams extracting tables, formulas and text from varied single-page documents at scale.
Alternatives
PaddleOCR-VL-1.6: common questions
- Is PaddleOCR-VL-1.6 worth it?
- Top-of-table document parsing in under a billion parameters, as long as you accept Baidu's toolchain. We rate it 4.4 out of 5.
- How much does PaddleOCR-VL-1.6 cost?
- Free under Apache 2.0. No first-party hosted price published.
- What are the downsides of PaddleOCR-VL-1.6?
- Headline scores are self-reported; full-page parsing needs the PaddlePaddle stack; vLLM serving needs a compute-capability 8.0+ GPU; no first-party hosted API or published throughput.
- Who is PaddleOCR-VL-1.6 best for?
- Teams extracting tables, formulas and text from varied single-page documents at scale.
- How much does PaddleOCR-VL-1.6 cost for a team or in production?
- Weights are free and about 1.9 GB in BF16, so a single RTX 30-series or A10 card serves it; no first-party pages-per-second figure or hosted price is published.
- Is PaddleOCR-VL-1.6 safe for company data?
- Self-hosted only, so documents stay on your hardware. Apache 2.0 permits commercial use and fine-tuning with no revenue cap.
- What are the best PaddleOCR-VL-1.6 alternatives?
- The closest alternatives we have reviewed are Unlimited-OCR and DeepSeek-OCR 2.
Answers checked Oct 22, 2026.
Make PaddleOCR-VL-1.6? Show this rating on your site
