PaddleOCR-VL 1.6, online

What changed in PaddleOCR-VL 1.6

PaddlePaddle's model card describes 1.6 as a compact 1.0B-parameter parser built on PaddleOCR-VL 1.5. Its region-aware data process targets weak document regions and improves supervision before progressive post-training.

PaddlePaddle reports 96.33% on OmniDocBench v1.6 and says the architecture remains compatible with 1.5. Those are upstream results. They are not measurements of this browser conversion, where page content, device memory, and the selected precision still decide practical quality and speed.

INT8 and FP32 in the browser

• INT8 decoder, FP32 vision and embeddings — about 2,434 MB. The smaller 1.6 package keeps the visual path at full precision and quantizes the decoder. The app guidance is 6+ GB memory.
• FP32 — about 3,466 MB. Full exported precision for maximum numerical fidelity. The app guidance is 8+ GB memory.

These are approximate browser download sizes, not the size of PaddlePaddle's original BF16 checkpoint. The memory values are compatibility guidance, not measured peaks.

In our tests

In our tests, both versions produced useful document text:

• INT8 recovered a complete model comparison table with usable structure. It also read a page containing mixed English and Japanese, although it skipped the opening paragraph.
• FP32 recovered the complete table content and the mixed-language prose with one Japanese substitution. The table came through as readable lines rather than a Markdown table.

These are targeted checks on document pages, not a general accuracy claim.

Version names are not interchangeable

PaddleOCR-VL 1.6 does not replace the existing PaddleOCR-VL 1.5 options. Version 1.6 downloads separately, so a previously downloaded 1.5 model is left alone.

PaddleOCR-VL 2 is a separate model generation, not another name for 1.6. Both earlier generations remain available.

Which model should you pick

Use PaddleOCR-VL 1.6 when you want promptable document parsing with text, tables, formulas, charts, seals, and page structure. Start with INT8 unless you have the bandwidth and memory for FP32.

GLM-OCR offers four WebGPU precision choices and broader explicit language coverage. Granite-Docling is available in five WebGPU variants for structured document conversion.

Open the model chooser to compare the available cards without removing your existing PaddleOCR-VL 1.5 selection.

Related reading