PaddleOCR-VL-1.6
Open Weights
Baidu
Baidu's compact document parsing vision-language model, served by the vendor as PaddleOCR-VL-1.6-0.9B. It handles text, tables, formulas, charts, seals and text spotting through task-specific prompts, and the vendor reports a 96.33% overall score on OmniDocBench v1.6. Version 1.6 adds a region-aware data pipeline that targets weak regions found in PaddleOCR-VL-1.5 plus a progressive post-training recipe with reinforcement learning. The architecture is unchanged from 1.5, so it is a drop-in replacement.
Strengths
- Vendor-reported 96.33% overall on OmniDocBench v1.6, described as a new state-of-the-art score
- One compact model, served by the vendor as PaddleOCR-VL-1.6-0.9B, covering text, tables, formulas, charts, seals and text spotting
- Task-switchable prompts for OCR, Table Recognition, Formula Recognition, Chart Recognition, Spotting and Seal Recognition
- Specific gains claimed on Chinese ancient documents and Chinese rare characters, alongside seal, spotting and chart recognition
- Architecture is fully compatible with PaddleOCR-VL-1.5, giving what the vendor calls zero adaptation cost, plug-and-play replacement
- Three deployment paths: the PaddleOCR CLI and Python API, a prebuilt vLLM inference server image, or transformers v5 directly
- Apache 2.0 licensed with weights on both Hugging Face and ModelScope
Caveats
- Almost every published figure is an image, not text: the OmniDocBench v1.6 and Real5-OmniDocBench comparison tables are PNGs, so only the 96.33% headline is citable and no per-category scores are recorded here
- The vendor evaluated its own model, Gemini-3 Pro and Qwen3-VL-235B-A22B-Instruct independently rather than taking those rows from the OmniDocBench leaderboard, so the published comparison is not like-for-like
- Real5-OmniDocBench is a benchmark the vendor built itself from OmniDocBench v1.5, so the 'new SOTA records' there are self-reported on self-made data
- The transformers path supports only element-level recognition and text spotting; page-level document parsing requires the official PaddleOCR pipeline, which the vendor recommends instead
- Deployment is constrained: PaddlePaddle 3.2.1 or above plus 'a special version of safetensors' is required, and macOS users are told to use Docker
- No context window, output limit or training cutoff is stated on the model card
- Vendor
- Baidu
- License
- Apache-2.0
- Release Date
- 2026-05-28
- Modalities
- imagetext
| Benchmark | Score | Max |
|---|---|---|
| OmniDocBench v1.6 (overall) | 96.3 | 100.0 |
Capabilities
Vision
Audio
Video
Tool Use
Resources
Pricing
Open weights under Apache 2.0, free to self-host through the PaddleOCR CLI and Python API, a vLLM inference server image, or transformers v5. No vendor-hosted pricing is published on the model card.