Active parameter parity does not establish OCR or vision parity

Compare multimodal models using their image encoder, visual token budget, exact extraction metrics and acceptance behavior, not active language parameter counts alone. A broad visual reasoning advantage is a useful prior but cannot be converted into an exact identifier error reduction. This matters when one shared model must handle both photographed documents and scene text. Hold preprocessing constant for a model comparison, then separately tune each model's quality/latency frontier; evaluate the actual quantized deployment checkpoint.

Evidence/example — 2026-09-22: Google's Gemma 4 model card reports a roughly 550M vision encoder for 26B A4B versus 150M for E4B, OmniDocBench 1.5 average edit distance 0.149 versus 0.181, and MMMU Pro 73.8 versus 52.6. Source: https://huggingface.co/google/gemma-4-26B-A4B-it . Google recommends higher visual token budgets for OCR and small text: https://huggingface.co/google/gemma-4-E4B-it . Inference: 26B A4B is the stronger general vision candidate despite similar active/effective language parameter labels. Limits: these are vendor-reported benchmarks, not a matched quantized deployment test; neither document edit distance nor MMMU establishes exact vehicle-identifier correctness, document acceptance calibration, or image-request latency. No private images or customer results are included.