PaddleOCR-VL
A 0.9B document-parsing VLM that topped the global OCR leaderboard — 109 languages, Apache 2.0.
About
PaddleOCR-VL is Baidu PaddlePaddle team's OCR-specialized vision-language model, released October 16, 2025. At just 0.9B parameters — a NaViT-style dynamic-resolution visual encoder on an ERNIE-4.5-0.3B language backbone — it is purpose-built for document parsing: text, tables, formulas, charts, seals, and handwriting across 109 languages, outputting structured Markdown and JSON. At launch it scored 92.56 on OmniDocBench v1.5 (provider-reported), ranking #1 globally ahead of DeepSeek-OCR, and it held Hugging Face's trending-models top spot for five straight days.
The family kept moving: PaddleOCR-VL-1.5 (January 2026, 94.5%, 111 languages) and 1.6 (May 2026, 96.3%) are drop-in upgrades on the same 0.9B architecture. It runs in Transformers and vLLM, with official GGUF, ONNX, and MLX community builds, plus an MCP server for Claude Desktop integration. The official site doubled its free parsing quota to 20,000 pages per day.
Specs
Context window
Pricing
Free quota
API access
Weights
Released
Benchmarks
Provider-reported: OmniDocBench v1.5 92.56 (#1 at launch, ahead of DeepSeek-OCR); v1.5 refresh 94.5%; v1.6 96.3%.
Independent: 5 consecutive days atop Hugging Face trending models.
Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.
Use cases
Trust score — 8/10
Model collection verified live on Hugging Face (HTTP 200); Apache 2.0 license; Transformers/vLLM integration confirmed in upstream repos; launch ranking covered by independent press. Docked only for: headline benchmark numbers are provider-run — though the community adoption (MinerU, RAGFlow, Cherry Studio integrations) is independently observable.