datalab-to/chandra-ocr-2 quantized with oQ (oMLX v0.5.3) at level 2.5, data-driven mixed precision: oQ measures each layer's quantization error through calibration and allocates bits where the measurement says they matter, rather than by a fixed per-tensor rule.

Table-column fidelity: PRESERVED. On the OCR fidelity check this build reads each config-table column into its own field.

Vision weights are kept fp16; only the language tower is quantized.