GLM-OCR, online

What GLM-OCR is

Z.ai describes GLM-OCR as a 0.9B-parameter multimodal OCR model built around a CogViT visual encoder, a cross-modal connector, and a GLM-0.5B decoder. Its task prompts cover text, formula, and table recognition; the upstream SDK adds PP-DocLayout-V3 for page layout analysis.

The upstream model card reports 94.62 on OmniDocBench v1.5. That is Z.ai's benchmark result, not a chunks.md browser measurement. Browser speed and memory depend on the GPU, driver, page, prompt, and precision selected here.

Four WebGPU browser variants

chunks.md uses this GLM-OCR ONNX build. Approximate downloads are:

• Q4F16 — 628 MB. Four-bit weights with FP16 compute. Smallest practical browser option; start here on a 4 GB-class GPU.
• INT8 — 1,249 MB. Eight-bit stored weights. The WebGPU graph expands quantized values during compute, so the compatibility guidance is 8+ GB VRAM, not half the FP32 requirement.
• FP16 — 2,119 MB. No integer weight quantization; intended for GPUs with reliable FP16 support. Guidance: 6+ GB VRAM.
• FP32 — 4,231 MB. Maximum exported precision and the heaviest download. Guidance: 10+ GB VRAM.

These are WebGPU-only options; there is no silent CPU/WASM fallback. The VRAM figures are compatibility guidance, not measured peaks. A device above the stated floor can still hit browser or driver allocation limits.

In our tests

In our tests, all four variants recovered a complete model comparison table with usable HTML structure and one model-name substitution. None produced repeated text or truncation, and higher precision did not visibly improve that page.

Only Q4F16 was checked on a Japanese news page. It returned the full page in readable structure with several character and word substitutions. Treat the other three variants as tested on English structured content only. These are targeted checks on document pages, not a general accuracy claim.

When to choose it

GLM-OCR is the broad multilingual choice when a page mixes ordinary text with tables, formulas, code, seals, or structured fields. The official model lists eight OCR languages: English, Chinese, Japanese, Korean, French, German, Spanish, and Russian.

Choose Q4F16 for the shortest download, FP16 when quality matters and memory permits, and FP32 only when the extra bandwidth and allocation risk are justified. INT8 sits between them in download size, but its current WebGPU compute path makes memory behavior less intuitive than the label suggests.

Compare the other document models

PaddleOCR-VL 1.6 is the direct upgrade path for PaddleOCR-VL users. Granite-Docling is available when structured DocTags matter more than broad multilingual recognition.

Open the model chooser, select the precision that fits your device, and expect the first run to download and cache that variant locally.

Related reading