PaddleOCR-VL
Baidu PaddlePaddle's 0.9B Apache 2.0 document parser: 96.33% on OmniDocBench v1.6 as reported, with tables, formulas, charts and seals.
In short. PaddleOCR-VL is a very small AI model that reads documents. It finds the layout of a page and turns text, tables, formulas, charts and stamps into structured output. At under 1 billion parameters it runs on modest hardware, and the Apache 2.0 license allows commercial use.
Summary
PaddleOCR-VL-1.6 was released on 28 May 2026 by the PaddlePaddle team at Baidu, as an upgrade of PaddleOCR-VL-1.5. The model is described as a 0.9-billion-parameter document parser; the checkpoint holds about 0.96 billion BF16 parameters, which is 1.92 GB, and uses the Apache 2.0 license. The architecture is fully compatible with version 1.5, so existing deployments can swap it in without changes.
According to the technical report, the improvement comes from two training changes: a region-aware data process that finds the places where the previous model was unstable, adds targeted training data there and cleans up unreliable labels, followed by a staged post-training recipe with data selection and reinforcement learning. The model card lists better recognition of tables, ancient Chinese documents and rare Chinese characters, plus gains in seal recognition, text spotting and chart recognition.
PaddleOCR reports 96.33% on OmniDocBench v1.6, which it presents as a new best score, and says it sets records on OmniDocBench v1.5 and on Real5-OmniDocBench, a set of scanned, warped, photographed, badly lit and skewed pages that the PaddlePaddle team built from the v1.5 data. Competitor figures come from the OmniDocBench leaderboard, except Gemini 3 Pro, Qwen3-VL-235B-A22B-Instruct and PaddleOCR's own model, which it evaluated itself. All of this is developer-reported, and the v1.6 score is not comparable with the 94.62 that GLM-OCR reports on V1.5.
The recommended route is PaddleOCR's document-parser pipeline, which needs PaddlePaddle 3.2.1 or later, optionally with a vLLM-based inference server in Docker. A Transformers route (version 5 or later) exists but handles only element-level recognition and text spotting, and Baidu also publishes a GGUF build.
Good for
- Full document parsing with layout, tables, formulas, charts and seals on modest hardware
- Chinese-heavy documents, including ancient texts and rare characters
- Replacing PaddleOCR-VL-1.5 without code changes
- Commercial pipelines that want the Apache 2.0 license
Watch out for
- The scores are PaddleOCR's own and use OmniDocBench v1.6, so they cannot be compared directly with results on v1.5
- The default route needs the PaddlePaddle framework 3.2.1 or later and a special version of safetensors
- The Transformers route covers only element-level recognition and text spotting, and macOS users are told to use Docker
- It is a document parser, not a general vision-language model
Performance claims come from the developer and have not been independently verified.
How much memory does it need?
Estimates for the PaddleOCR-VL-1.6 variant. A mixture-of-experts model must hold all of its experts in memory even though only some are used for each token, so total parameters decide the memory and active parameters decide the speed.
| Variants | Parameters | BF16 | FP8 / INT8 | 4-bit (Q4) | Fits on |
|---|---|---|---|---|---|
| PaddleOCR-VL-1.6 BF16 | 960M | 2.21 GB | 1.1 GB | 0.662 GB | Laptop or phone |
Estimates: weights plus about 15% overhead. Real usage grows with context length and depends on the software.
The license in plain words
One of the most widely used permissive licenses. You may use, change and sell products built on the model with only light conditions.
See also
- GLM-OCR vs PaddleOCR-VL: which small document parser should you use?
- Best open-source OCR and document models in 2026
- Eleven open OCR models in a year: small and hard to compare
Frequently asked questions
- What is PaddleOCR-VL?
- PaddleOCR-VL is a very small AI model that reads documents. It finds the layout of a page and turns text, tables, formulas, charts and stamps into structured output. At under 1 billion parameters it runs on modest hardware, and the Apache 2.0 license allows commercial use.
- Can I use PaddleOCR-VL commercially?
- Yes. PaddleOCR-VL is released under Apache 2.0, a permissive license. You can use it commercially if you follow the standard conditions, such as keeping the license notice.
- Is PaddleOCR-VL open source?
- PaddleOCR-VL is released under Apache 2.0, a license approved by the Open Source Initiative. Training data and training code may not be public, so it is best described as open weights under an open-source license.
- Can I run PaddleOCR-VL on my own computer?
- Yes. The smallest variant, PaddleOCR-VL-1.6, needs about 0.662 GB of memory at 4-bit precision, which fits a modern laptop or a mid-range graphics card.
- When was PaddleOCR-VL released?
- PaddleOCR-VL was released by Baidu on May 28, 2026.
Official links
Sources
Specifications come from these primary sources. Benchmark claims, where mentioned, are reported by the developer.
- PaddleOCR-VL-1.6 model card (Hugging Face) huggingface.co
- PaddleOCR-VL-1.6 technical report arxiv.org
- PaddleOCR repository (GitHub) github.com
Release date and facts last verified November 11, 2026.
Related models
| Model | Released | Parameters | Context | License | Smallest variant needs |
|---|---|---|---|---|---|
| GLM-OCR Z.ai (Zhipu AI) | 900M | – | MIT (Permissive) | 0.621 GB Laptop or phone | |
| Unlimited-OCR Baidu | 3.34B | – | MIT (Permissive) | 2.3 GB Laptop or phone | |
| DeepSeek-OCR 2 DeepSeek | 3.39B | – | Apache 2.0 (Permissive) | 2.34 GB Laptop or phone | |
| Mirage Cactus Compute | 51M | – | Apache 2.0 (Permissive) | 0.0354 GB Laptop or phone |