- New SOTA: PaddleOCR-VL-1.6 scores 96.33% on OmniDocBench v1.6, beating all open-source and closed-source models including 235B+ systems.
- Tiny but mighty: The 0.9B model outperforms Gemini 3 Pro and Qwen3-VL-235B on real-world robustness benchmarks.
- Key innovation: An "Under-Optimized Region" data engine finds and fixes unstable, sparse, and mislabeled training examples instead of adding more data.
- Three-stage training: Progressive CPT → SFT → GRPO reinforcement learning pipeline; each stage contributes measurable gains.
- Free and local: Apache 2.0 license, available as GGUF on Hugging Face, runs via llama.cpp, Ollama, or LM Studio.
- Plug-and-play: Fully architecture-compatible with PaddleOCR-VL-1.5; zero migration cost for existing users.
PaddleOCR-VL-1.6 just set a new state-of-the-art on document parsing benchmarks, and the headline number almost doesn't make sense at first glance: a 0.9-billion-parameter model outperforming systems with 235 billion parameters. Baidu's PaddlePaddle team didn't get there by scaling up. They got there by being surgical about what was broken.
The problem with "just add more data"
Document parsing has become a central interface between unstructured documents and LLM applications. Modern document systems are expected to recover not only plain text, but also layout regions, reading order, mathematical formulas, tables, charts, seals, and spatially grounded text instances -- and convert all of it into Markdown, JSON, or other machine-readable formats for downstream indexing, retrieval, and reasoning.
The predecessor, PaddleOCR-VL-1.5, was already a strong baseline. But the team noticed something important: the remaining errors concentrated in under-optimized regions where model behavior was unstable, data coverage was sparse, or supervision was unreliable. Rather than dumping more training data into the pipeline, they built a system to find and fix exactly those weak spots.
Diagnosing what's actually broken
The core innovation is what the paper calls an Under-Optimized Region (UOR) driven data engine. It diagnoses three distinct failure modes in the previous model:
- Boundary-Fragile Regions: samples where small pixel-level shifts or visual distortions cause large changes in model output. The team evaluated every training sample across 8 late-stage checkpoints and 16 semantic-invariant distortions -- generating 8,128 pairwise comparison scores per sample -- to find the top 1% most unstable examples.
- Coverage-Sparse Regions: areas of the training distribution that are underrepresented. Think ancient manuscripts, rare characters, and industrial tables. The engine clusters samples by visual-semantic similarity and finds the small, weakly-connected outlier groups that dominant distributions tend to absorb.
This story is for Pro members
You've reached the end of the free preview. Upgrade to AlphaSignal Pro to read the full article - and everything else behind the paywall.
Comments
