PaddleOCR-VL-0.9B: Multilingual Document Parsing

• PaddleOCR-VL-0.9B is an ultra-compact vision-language model designed for multilingual document parsing across 109 languages, recognizing text, tables, formulas, and charts.
• The model integrates a NaViT-style dynamic resolution visual encoder with a lightweight MLP projector and ERNIE-4.5-0.3B language model to maintain spatial fidelity and resource efficiency.
• It achieves state-of-the-art performance on benchmarks with improvements in throughput, memory usage, and accuracy in element extraction for real-world document automation.

PaddleOCR-VL-0.9B is an ultra-compact vision-LLM (VLM) for multilingual document parsing, designed to efficiently recognize and structurally extract diverse elements—including text, tables, formulas, and charts—across 109 languages. It integrates a NaViT-style dynamic resolution visual encoder with the ERNIE-4.5-0.3B LLM and a lightweight MLP projector, forming the core of the PaddleOCR-VL architecture. The model is optimized for resource efficiency and inference speed, outperforms prior solutions on key public and in-house benchmarks, and is suitable for deployment in real-world systems spanning document digitization, business process automation, and multilingual information retrieval (Cui et al., 16 Oct 2025).

1. Model Architecture

PaddleOCR-VL-0.9B comprises three principal components:

• NaViT-style Dynamic Resolution Visual Encoder: This module accepts native-resolution document images, avoiding resizing-induced distortions and preserving spatial fidelity in text-dense regions. It outputs high-resolution visual feature vectors suitable for parsing complex page layouts.
• Randomly Initialized 2-Layer MLP Projector: Bridges the visual encoder and LLM by mapping the feature vectors into LLM-compatible embeddings. With a merge size parameter (e.g., 2), it ensures compact token representations. The projection function is explicitly described as:

$$E = \text{GELU}(W_2 \cdot (W_1 \cdot V + b_1) + b_2)$$

where $V$ are visual features and $W_1, W_2, b_1, b_2$ are learnable parameters.

• ERNIE-4.5-0.3B LLM: A compact autoregressive LLM leveraging enhanced 3D-relative position encoding (3D-RoPE) to maintain spatial relationships crucial for correct reading order and element segmentation. It generates structured output (e.g., Markdown, JSON) reflecting the page's hierarchical organization.

This design enables efficient end-to-end element-level recognition with explicit attention to global layout structure.

2. Multilingual and Multi-Element Capabilities

The model's broad language support (109 languages) is achieved by training on over 30 million heterogeneous samples spanning public, synthetic, and in-house sources. These cover:

• Varied writing systems: Latin, Cyrillic, Devanagari, Arabic, etc.
• Diverse domains and formats: academic papers, financial documents, newspapers, ancient texts, handwritten manuscripts.
• Complex layouts and mixed-content scenarios: vertical text, script ambiguity, and densely intermixed languages.

Element recognition extends beyond plain text to include:

• Detailed table structure recovery
• Formula recognition with high Character Detection Matching (CDM) scores (e.g., CDM = 0.9453 on OmniDocBench-formula-block)
• Chart extraction (RMS-F1 = 0.8440) facilitating downstream extraction from visual plots

Explicit modeling of both character-level and layout-level signals allows PaddleOCR-VL-0.9B to avoid typical error cases seen in specialist or monolingual OCR systems.

3. Benchmarking and Performance Metrics

Extensive evaluations confirm the model's state-of-the-art performance:

• OmniDocBench v1.5: Overall score of 92.56; Text Edit Distance = 0.035; Formula CDM = 91.43; leading Table-TEDS and Reading Order Edit Distance performance.
• In-house Element-Level OCR Tasks: Lowest averaged edit distances across both multilingual and element-specific metrics.
• Inference Efficiency: On an NVIDIA A100 GPU using the vLLM backend:
  • Throughput: 1.224 pages/sec, 1881 tokens/sec
  • VRAM usage: ~43.7GB (with 40% lower memory consumption than some competitive baselines)
  • Page throughput improved by 15.8% over contemporary models like MinerU2.5 and dots.ocr.

The inference pipeline is highly parallelized, with asynchronous loading, dynamic batching, and decoupled layout analysis and VLM inference, further reducing latency and maximizing hardware utilization.

4. Resource Optimization and Scalability

PaddleOCR-VL-0.9B’s architecture achieves notable resource efficiency:

• Parameter Size: 0.9B parameters, categorizing it as ultra-compact within the VLM landscape
• Dynamic Resolution Processing: Avoids excessive tiling, maintaining detail without computational redundancy
• Memory and Compute Management: Efficient GPU utilization enables deployment in resource-constrained environments or high-throughput cloud services
• Pipeline Design: Multi-threading and asynchronous item queue allow maximal exploitation of hardware parallelism

This enables scaling to enterprise-level document streams and integration into high-traffic scenarios without the infrastructure burden typically imposed by billion-parameter VLMs.

5. Real-World Deployment and Application Scenarios

PaddleOCR-VL-0.9B is applicable to a broad array of document-centric tasks:

• Document Digitization and Archiving: High precision parsing of legal manuscripts, scientific papers, and historical records.
• Business Process Automation: Automated extraction of structured data from financial reports, invoices, and forms, including complex tables and charts.
• Multilingual Content Extraction: Reliable analysis across mixed-language corpora for multinational or governmental organizations.
• Retrieval-Augmented Generation (RAG) Pipelines: Provision of structured, contextualized document representations for downstream LLMs.
• Cloud Inference Services: Efficient integration with vLLM or FastDeploy backends for real-time, large-scale serving.

Its robustness to layout and language heterogeneity positions it as a practical backbone for advanced document understanding systems.

6. Technical Innovations and Integration Context

The model integrates and extends several key techniques:

• NaViT-style dynamic resolution encoding for native image feature acquisition
• MLP projection with nonlinear activation and merge-size compression for resource-optimized embedding
• Autoregressive decoding with spatially-aware token generation via 3D-RoPE
• End-to-end element-level recognition as opposed to piecemeal (text-only) approaches
• Plug-and-play role within broader PaddleOCR document parsing pipelines, compatible with upstream layout analysis and downstream information extraction modules

This positions PaddleOCR-VL-0.9B as a critical evolution in the PaddleOCR family, superceding prior pipeline systems through a unified, compact vision-language approach.

7. Summary Table: Model Features and Benchmark Metrics

Feature	Details	Benchmark Performance
Parameters	0.9B	15.8% ↑ throughput vs. MinerU2.5
Visual Encoder	NaViT-style, dynamic resolution	Text Edit Distance = 0.035 (OmniDocBench)
LLM	ERNIE-4.5-0.3B w/ 3D-RoPE	Formula CDM = 0.9453 (OmniDocBench)
Element Recognition	Text, tables, formulas, charts	RMS-F1 (chart) = 0.8440
Multilingual Coverage	109 languages	Leading scores on multi-lingual and structured benchmarks
Page Throughput (@A100/vLLM)	1.224 pages/sec; 43.7GB VRAM	40% ↓ memory usage vs. competitive baselines

The above table summarizes the architectural components, principal features, and representative benchmark improvements as reported (Cui et al., 16 Oct 2025).

---

PaddleOCR-VL-0.9B constitutes a major advance in compact, high-throughput multilingual document parsing, combining dynamic visual encoding, resource-efficient projection, and autoregressive language modeling with explicit support for complex layouts and structured elements. Its empirical performance and deployment characteristics position it as a leading solution for both academic research and industrial-scale document automation.