Skip to content
Bud's Wiki

PaddleOCR-VL: Boosting Multilingual Document Parsing via a 0.9B Ultra-Compact Vision-Language Model

TL;DR

Section titled “TL;DR”

PaddleOCR-VL is a 0.9B-parameter document-parsing VLM from Baidu that pairs a NaViT-style dynamic-resolution visual encoder with the ERNIE-4.5-0.3B language model. The pitch is SOTA on page-level document parsing (OmniDocBench v1.0 / v1.5) and on element-level recognition (text, tables, formulas, charts) while staying small enough to be deployable, with 109-language coverage including scripts like Cyrillic, Arabic, Devanagari, and Thai. The model and weights are Apache-2.0 on HuggingFace, with transformers + vLLM + flash-attn-2 inference paths.

Key claims

Section titled “Key claims”
  • The full system is a 0.9B VLM composed of a NaViT-style dynamic-resolution visual encoder plus an ERNIE-4.5-0.3B language model, targeted at element recognition and document parsing [Introduction; Core Features].
  • Claims SOTA on OmniDocBench v1.5 across overall, text, formula, tables, and reading-order metrics [Performance §Page-Level §1].
  • Claims SOTA on “almost all metrics” of overall, text, formula, tables, and reading order on OmniDocBench v1.0 [Performance §Page-Level §2].
  • Supports 109 languages including Chinese, English, Japanese, Korean, Latin, Russian (Cyrillic), Arabic, Hindi (Devanagari), and Thai [Core Features §Multilingual Support].
  • Element-level results: lowest edit distance across all scripts on in-house OCR and best performance per category on in-house Formula (simple/complex print, camera scan, handwritten) [Performance §Element-level §1, §3].
  • On in-house Chart (11 categories including bar-line hybrid, pie, 100% stacked bar, area, bubble, scatterplot, etc.), the 0.9B model is reported to outperform expert OCR VLMs and “some 72B-level multimodal language models” [Performance §Element-level §4].
  • The HuggingFace card pegs OmniDocBench v1.5 overall at 94.5% [tweet announcement; HF model card].

Method

Section titled “Method”

PaddleOCR-VL plugs a NaViT-style dynamic-resolution vision encoder into ERNIE-4.5-0.3B. The vision branch handles variable input resolutions natively (NaViT-style “native resolution” packing) which matters for documents with mixed densities — paragraphs vs. tables vs. dense formulas — without forcing a fixed crop. The language model is the small ERNIE-4.5-0.3B, giving the entire VLM a 0.9B-parameter footprint. At inference time the model is exposed both as a full document-parsing pipeline (paddleocr doc_parser, returning JSON + markdown) and as a plain transformers element-recognizer that takes one of four prompt modes — OCR:, Table Recognition:, Formula Recognition:, Chart Recognition: — and returns the corresponding decoded structure for the input image. Production serving goes through vLLM with flash-attention-2; the repo also ships a Docker image for a PaddleOCR-managed vLLM endpoint that the CLI can hit.

Results

Section titled “Results”

Headline OmniDocBench v1.5 overall score is 94.5% per the HF/tweet card; the model is reported SOTA across all five major v1.5 subscores and across “almost all” v1.0 subscores. Element-level results are reported on a mix of public OmniDocBench-OCR-block and in-house benchmarks: lowest edit distance across all scripts on in-house OCR, best per-category numbers on in-house Tables (Chinese, English, mixed, full/partial/no-border, book/manual, list, academic, merged-cell, low-quality, watermarked), best per-category on in-house Formulas, and outperforming “some 72B-level multimodal language models” on the 11-category in-house Chart benchmark — at 0.9B total parameters. The card credits MinerU and OmniDocBench for the public-side numbers and notes internal evaluations on top. Concrete inference modes: transformers (element-level only), vLLM (page-level), and a PaddleOCR-supplied vLLM Docker server.

Why it’s interesting

Section titled “Why it’s interesting”

Second specialist OCR-VLM landing in the same week as DeepSeek-OCR-2 (per Haoxiang’s note) — a pattern of “shrink a general VLM into a 0.5-1B specialist” rather than scaling up a general MLLM for documents. The NaViT-style dynamic-resolution encoder + small LM recipe is shared with the broader compact-VLM trend (Qwen3-VL-Embedding, the ERNIE-Image-Aes line, MiniCPM-V), and the chart-understanding claim — beating 72B generalists at 0.9B with a task-specialized decoder — is the kind of efficiency-frontier datapoint worth tracking against in-house dense-captioning and multimodal-reasoning needs. Also notable: Apache-2.0 weights from Baidu, which makes it actually usable as a captioner / pre-processor in a Luma data pipeline if document OCR ever shows up as a bottleneck.

See also

Section titled “See also”