GLM-OCR, FireRed-OCR hit CPU limits

This is less an OCR problem than a table-geometry problem: once merged cells and handwriting show up, the model is being asked to infer structure, not just read text. The thread points to the right instinct, which is a deterministic preprocessing layer plus a smaller model pass, not one giant prompt. GLM-OCR is the closest fit among the named tools because the official repo says it is a 0.9B multimodal OCR model with Markdown and JSON output, but the self-hosted path still centers on heavier serving stacks rather than a pure CPU-only laptop workflow. FireRed-OCR looks impressive on document benchmarks and structural integrity, yet its public quick start uses `device_map="auto"` and `bfloat16`, which suggests it expects accelerator-backed inference. HTML or Markdown intermediates help, but only if they are chunked by logical regions or table bands first; otherwise the context window just becomes the next bottleneck. The Excel side should skip vision entirely: unmerge cells, flatten the headers, and extract schema metadata directly from sheet structure. The Reddit replies land on a pragmatic tradeoff: rent GPU time or use a stronger multimodal model for the pathological cases, then keep the CPU-local stack for preprocessing and validation.