GLM-OCR (9B) - Local OCR Test | OCR, Document Extraction, Table Recognition

GM-OCR is a two-stage OCR system that combines document layout analysis with character-level recognition, and it’s drawing attention because it delivers strong real-world extraction quality while staying small enough to run locally. The model’s practical appeal hinges on its size—about 0.9B parameters—and its MIT license, which makes it easier to deploy in private or offline document pipelines. In benchmarks it’s positioned as especially effective for complex tables, code-like content, and visual elements such as figures and charts, and the transcript’s local test aims to validate that promise outside of curated benchmark settings.

The testing setup runs GM-OCR in a Google Colab notebook using a T4 GPU with roughly 16 GB of VRAM. The model weights are about 2.2 GB, and the saved tensors file is reported around 2.7 GB—small enough to fit comfortably on common consumer GPUs. The workflow uses the Hugging Face Transformers library (noted as requiring a latest version, 5.1 in the transcript), with an auto processor for image preprocessing and an auto model for image-to-text generation. The model is loaded in full 16-bit floating-point precision and moved entirely onto the GPU for inference. Quantized variants are also mentioned as available via OA and CP, which could speed up inference further.

GM-OCR’s inference is prompt-driven and supports a limited set of prompt types: general text recognition, table recognition, and custom extraction via a schema that returns results in JSON. A single recognition pass on a document takes about 53 seconds initially, with subsequent runs faster—an important operational detail for anyone building batch extraction.

On document quality, the results are mixed but often impressive. For a document with a title and formatted text, the extracted text is described as nearly perfect across a manual check of roughly 60–70% of the content, with only minor punctuation issues. Tables, however, don’t always come out as clean Markdown tables; instead, the transcript shows table content appearing in less structured text form. In a large “one giant table” financial report, the extraction is largely accurate for values and spacing, but some column headers (including “2025” and “2025”) are missing or misplaced, suggesting table reconstruction still has edge cases.

A receipt test—described as skewed but well laid out—produces strong results for key fields like unit, address, telephone number, date, time, and even card numbers. The most compelling capability is custom field extraction: by specifying fields such as address, telephone, date, time, order, and total, the model returns JSON with values that match prior extraction, though the total may omit the currency symbol.

Finally, a fake ID card shows mostly correct extraction but includes small errors, such as the date of birth being off by a couple of years (1908 vs. 1906). Overall, the transcript frames GM-OCR as a strong performer for a relatively small model, with practical deployment options ranging from full-precision local runs to quantized inference for speed.