Ollama Community Reports GLM-OCR Table Recognition Regression; Upgrade May Produce Plain Text or Repetitive Output

A report of degraded GLM-OCR table recognition appeared on Ollama’s GitHub on October 5. The user said that after upgrading to 0.35.1, the same document image and Table Recognition: prompt could produce only plain text or enter a repetitive generation loop, eventually ending with HTTP 500. The case concerns the structure and stability of document-parsing output, making it directly relevant to workflows that rely on table data. Community report

The test environment was Windows 11 with an RTX 5060 Ti 16GB and the model library’s glm-ocr:latest bf16 model. The reporter ran versions 0.34.0 and 0.35.1 side by side, saying the older version produced complete HTML tables in 4 to 9 seconds, while some requests on the newer version took 1 to 5 minutes. On the reporter’s synthetic lab-report test set, the number of correct values fell from 197/197 to 113/197. These figures come from a single user’s tests and cannot be taken as representative of document accuracy in general. Version comparison

The table prompt is documented by the official sources. The Ollama model page lists Table Recognition: as a usage pattern, and the Z.ai model card also lists text, formula, and table recognition among supported tasks. GLM-OCR combines a vision encoder, a multimodal connector, and a language decoder. The official full document-parsing SDK adds layout analysis and structured output, so results from direct model inference and the full parsing pipeline should be evaluated separately. Ollama usage guide, official model card

The engineering risk is that downstream systems may misalign fields if cell structure is missing, even when the model can still read the text. This could affect report imports or the quality of data used for retrieval. That is a possible consequence inferred from the case; there is currently no evidence that other deployments are broadly affected. When validating an upgrade, teams can check content accuracy, HTML parseability, row and column alignment, and timeout rates.

The reporter noted that both versions logged the same startup parameters and therefore suspected a change in the bundled llama-server build; this remains an unverified hypothesis. Reducing image size or adding a newline at the end of the prompt only helped occasionally. Next, maintainers would need to confirm whether they can reproduce the issue, identify the version difference, and verify the fix against the same images. Issue and temporary workarounds