PaddleOCR-VL-1.5: Comprehensive Analysis of the 0.9B SOTA Document Parsing Model

Introduction: A New Milestone in Document Parsing

In today's rapidly evolving AI landscape, document parsing technology has become a critical bridge connecting the physical and digital worlds. From academic papers to business contracts, from invoices to historical archives, massive volumes of document information urgently need to be accurately and efficiently digitized and structured. This goes beyond simple OCR (Optical Character Recognition) upgrades—it represents a comprehensive understanding of documents' deep semantics and structure.

On January 29, 2026, Baidu's PaddlePaddle team released PaddleOCR-VL-1.5, a multi-task Vision-Language Model (VLM) with only 0.9B (900 million) parameters that achieved 94.5% accuracy on the OmniDocBench v1.5 benchmark, setting a new State-of-the-Art (SOTA) record. Even more remarkably, this lightweight model outperformed massive general-purpose VLMs like Qwen3-VL-235B and Gemini-3 Pro in real-world robustness testing.

PaddleOCR-VL-1.5 represents more than just a performance improvement—it marks a paradigm shift in document parsing technology: from single text recognition to unified parsing of tables, formulas, charts, and seals; from recognition in ideal conditions to handling real-world challenges like scanning, skewing, warping, and screen photography. This marks the official entry of document parsing technology into a new era of "practicality" and "intelligence."

Core Highlights

1. Ultra-Lightweight Architecture with SOTA Performance

The most striking feature of PaddleOCR-VL-1.5 is its extreme parameter efficiency. With just 0.9B parameters, it achieves 94.5% accuracy on OmniDocBench v1.5, surpassing not only its predecessor PaddleOCR-VL-1.0 but also demonstrating remarkable advantages in comparisons with large general-purpose VLMs:

This parameter efficiency stems from PaddlePaddle team's deep understanding and careful design of document parsing tasks. The model employs a NaViT-style dynamic resolution visual encoder paired with the lightweight ERNIE-4.5-0.3B language model, maintaining high accuracy while dramatically reducing computational costs and deployment barriers.

2. Six Core Capabilities in a Unified Model

PaddleOCR-VL-1.5 is a true multi-task model, supporting six core capabilities within a single architecture:

This unified multi-task architecture not only simplifies deployment but, more importantly, enables knowledge sharing and collaborative optimization across different tasks, achieving better performance on each task.

3. Real-World Robustness: Real5-OmniDocBench

To evaluate the model's performance in real-world scenarios, the PaddlePaddle team constructed the Real5-OmniDocBench benchmark, covering five common physical distortion scenarios:

On this more practical test set, PaddleOCR-VL-1.5 achieved an overall accuracy of 92.05%, setting a new SOTA record. This means the model maintains stable high performance whether processing contract photos taken with a smartphone or historical documents processed by scanners.

Technical Architecture Deep Dive

PaddleOCR-VL-1.5 adopts an innovative two-stage architecture design that organically combines layout analysis and element recognition for end-to-end document parsing capabilities.

PP-DocLayoutV3: Unified Layout Analysis Engine

PP-DocLayoutV3 is the first stage of PaddleOCR-VL-1.5, responsible for document layout analysis. Unlike traditional rectangular detection boxes, PP-DocLayoutV3 introduces instance segmentation technology to predict precise pixel-level masks, crucial for handling skewed and warped documents.

Core Innovations:

PaddleOCR-VL-1.5-0.9B: Element-Level Recognition Model

The second stage, PaddleOCR-VL-1.5-0.9B, performs fine-grained recognition on elements obtained from layout analysis. The model inherits the lightweight architecture of PaddleOCR-VL-0.9B but with significantly expanded capabilities.

Architecture Components:

Training Strategy: Progressive training paradigm with three stages - pre-training (46M image-text pairs), post-training (5.6M instruction data), and reinforcement learning (GRPO optimization).

Performance Evaluation & Comparison

OmniDocBench v1.5: Comprehensive Leadership

OmniDocBench v1.5 is one of the most authoritative document parsing benchmarks, covering multiple element types including text, tables, formulas, and charts. PaddleOCR-VL-1.5 achieved 94.5% overall accuracy, setting new SOTA records across multiple sub-tasks:

Comparison with Competitors:

Real5-OmniDocBench: Real-World Performance

On Real5-OmniDocBench testing, PaddleOCR-VL-1.5 demonstrated exceptional robustness with an overall accuracy of 92.05%, surpassing all tested models including general-purpose VLMs with far larger parameter sizes.

Hardware Requirements & Deployment

Hardware Requirements

Recommended Configuration:

Minimum Configuration:

Supported Platforms: CUDA (NVIDIA GPU), ROCm (AMD GPU, Day 0 support), CPU inference

Deployment Options

1. Docker Deployment (Recommended)

docker run --rm --gpus all --network host \
    paddlepaddle/paddleocr-genai-vllm-server:latest-nvidia-gpu \
    paddleocr genai_server --model_name PaddleOCR-VL-1.5-0.9B

2. vLLM Accelerated Deployment

vllm serve PaddlePaddle/PaddleOCR-VL-1.5-0.9B \
    --host 0.0.0.0 --port 8080

3. Native PaddlePaddle Deployment

from paddleocr import PaddleOCRVL
pipeline = PaddleOCRVL()
output = pipeline.predict("document.pdf")

Use Cases & Best Practices

Typical Application Scenarios

Best Practice Recommendations

Conclusion

PaddleOCR-VL-1.5 represents a significant breakthrough in document parsing technology. With only 0.9B parameters, it achieves 94.5% SOTA accuracy and outperforms much larger general-purpose models in real-world robustness testing.

Core Advantages:

Related Links:

Link