- New SOTA: PaddleOCR-VL-1.6 scores 96.33% on OmniDocBench v1.6, beating all open and closed-source models.
- Tiny but mighty: Only 0.9B parameters — outperforms Qwen3-VL-235B and Gemini 3 Pro on document parsing benchmarks.
- Key innovation: Under-Optimized Region (UOR) Mining identifies and fixes boundary-fragile, coverage-sparse, and mislabeled training samples instead of scaling data blindly.
- Three-stage training: Progressive CPT (16.8M samples) → SFT (7.3M hard samples) → RL via GRPO (49K curated samples) pipeline drives the gains.
- Real-world robustness: Tops all five scenarios on Real5-OmniDocBench (scanning, warping, screen photography, illumination, skew) with 93.19% overall.
- Free and plug-and-play: Apache 2.0 license, fully compatible with PaddleOCR-VL-1.5, available via
paddleocrPython package and vLLM serving.
PaddleOCR-VL-1.6, Baidu's latest compact document parsing model, just set a new state-of-the-art on the authoritative OmniDocBench benchmark , and it did so with fewer than 1 billion parameters. That's not a typo. A 0.9B model is now outperforming 235B-parameter giants like Qwen3-VL and proprietary systems like Gemini 3 Pro on structured document understanding. The release is free, open-source under Apache 2.0, and available right now on Hugging Face.
The problem it's solving
Document parsing , turning scanned PDFs, photos of forms, or image-heavy reports into structured Markdown or JSON , is one of the most critical bottlenecks in building reliable RAG (retrieval-augmented generation) pipelines. The challenge isn't just reading text; it's faithfully reconstructing tables, formulas, charts, seals, and reading order from images that may be warped, poorly lit, or photographed at an angle. Most solutions either require massive models or sacrifice accuracy on edge cases like ancient manuscripts, rare characters, or complex nested tables.
Top systems increasingly encounter difficult regions that aren't fully addressed by increasing data volume or model size alone. Long-tail document layouts, rare scripts, dense formulas, complex tables, and noisy supervision can remain underrepresented or unreliable even when the overall training corpus is large. PaddleOCR-VL-1.6's entire design philosophy is a direct answer to this problem.
What's new in 1.6
PaddleOCR-VL-1.6 is an upgraded compact document parsing model built upon PaddleOCR-VL-1.5. It introduces a region-aware data optimization framework that identifies weak regions from the previous model, applies targeted enhancement to those regions, and improves the reliability of supervision signals. It further adopts a progressive post-training recipe based on curated data selection and reinforcement learning, pushing model performance to a higher level through staged optimization.
The improvements are concrete and measurable. Compared to 1.5:
- Overall score on OmniDocBench v1.6: 94.93% → 96.33%
- Table TEDS (tree edit distance similarity for tables): 91.67% → 94.76% (+3.09%)
- Table TEDS-Structure: 94.37% →
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
