Granite-Docling OCR online

Availability

Granite-Docling is available in model selection on chunks.md. It is WebGPU-only and has no CPU/WASM fallback. Start with Q4, which gave the best result in our checks, then choose another precision in Advanced settings if your device and workload justify it.

English is the primary language. Japanese, Arabic, and Chinese support is experimental, so use a multilingual OCR model when dependable recognition in those languages matters more than Docling's structured output.

What Granite-Docling is

IBM describes Granite-Docling 258M as an image-to-text model for efficient document conversion. It combines a SigLIP2 vision encoder with a Granite 165M language model and produces DocTags, which Docling can turn into markdown, HTML, and other document formats.

The model targets full-page OCR, layout, code, equations, charts, and table structure. English is its primary language; IBM labels Japanese, Arabic, and Chinese support experimental. The model and browser conversion use the Apache 2.0 license.

Plain OCR usually returns text. Granite-Docling returns a document representation, allowing headings, paragraphs, tables, code, formulas, and locations to survive conversion instead of arriving as one flattened block.

In our tests

In our tests, Q4 performed best. It produced a complete Markdown table with readable English and mostly recognizable entries, but several values were wrong. On a Japanese news page, the output collapsed into a repeated character instead of usable text.

Q8 produced numbered junk on the same English page. FP32 returned only a picture tag instead of the document text. Q4F16 and FP16 could not complete the evaluation because those exports hit a runtime incompatibility in our test path.

These were targeted checks on a small set of document pages, not a general accuracy claim. That is why Q4 is the recommended variant even though higher-precision exports are available for users who want to try them.

Browser variants

All five WebGPU-only exports are selectable:

• Q4F16 — 268.4 MB
• Q8 — 324.2 MB
• Q4 — 404.2 MB
• FP16 — 636.3 MB
• FP32 — 1.268 GB

Q4 is larger than Q4F16 because its upstream export keeps the embedding at full precision. Estimated requirements are 4+ GB VRAM / 8 GB RAM for Q4F16, Q8, and Q4; 6+ GB VRAM / 8 GB RAM for FP16; and 8+ GB VRAM / 12 GB RAM for FP32. These are compatibility estimates, not measured browser peaks.

When to choose another model

GLM-OCR is the stronger choice for multilingual text, tables, formulas, and structured extraction, with four WebGPU precision choices. PaddleOCR-VL 1.6 offers INT8 and FP32 while keeping PaddleOCR-VL 1.5 intact.

Open the model chooser and select Granite-Docling to run it locally.

Related reading