← All models

HPD-Parsing

Published Publisher Baidu

Overview

HPD-Parsing is a lightweight, approximately 1-billion-parameter vision-language model for high-throughput document parsing. The published model repository is PaddlePaddle/HPD-Parsing; this is a repository identifier, not a verified hosted API model ID. The model card lists the Apache 2.0 license. (huggingface.co)

Capabilities and uses

The model processes document-page images to recognize text and reconstruct page structure, reading order, formulas and tables. It is intended for document digitization and high-volume or batch parsing. Its hierarchical parallel decoding uses a main branch for page layout and concurrent branches for individual content blocks; Progressive Multi-Token Prediction reduces decoding steps. (arxiv.org)

The authors report a peak throughput of 4,752 tokens per second on public benchmarks. This is a benchmark result, not a guaranteed production rate. (arxiv.org)

Deployment and limitations

The official tutorial supports local inference and self-hosted OpenAI-compatible serving using a customized vLLM runtime. The documented setup requires an NVIDIA GPU with a driver supporting CUDA 12.8 or later; the prebuilt package targets Linux x86-64 and Python 3.10–3.13. Actual throughput and parsing accuracy depend on the workload and environment, and results should be checked for errors. (paddleocr.ai)

Release date

The official paper listing states that the first public release was July 21, 2026. The PaddleOCR repository announced availability on July 22, 2026. (github.com)

Pricing

As of November 16, 2026, an official HPD-Parsing-specific hosted price could not be verified on either Baidu's mainland China or international pricing pages. Input and output charges are therefore unverified; no CNY or USD amount is asserted. Self-hosting requires infrastructure, whose costs depend on the deployment. Other OCR products and models listed on those pages should not be assumed to use HPD-Parsing. (cloud.baidu.com)

Official sources