HPD-Parsing Achieves SoTA in Document Parsing

HPD-Parsing is a lightweight (1B), high-throughput document parsing model released by PaddlePaddle, based on the Hierarchical Parallel Decoding (HPD) paradigm.

Existing VLM-based parsers generate an entire page through single token-by-token autoregression, creating a bottleneck proportional to document length. HPD-Parsing addresses this by:

• A main layout branch that coordinates the global document structure and dynamically distributes region-specific content generation to parallel branches
• Progressive Multi-Token Prediction (P-MTP), which further shortens decoding steps within each branch
• Shared-prefix KV cache reuse to improve efficiency

Benchmark results:

• OmniDocBench v1.6 overall score 94.91% — the highest among end-to-end integrated parsers
• Maximum throughput of 4,752 TPS — 2.62× faster than the previous fastest parser, and 3.06× faster than its own autoregressive baseline

The model is publicly available on Hugging Face.