datalab-to/chandra-ocr-2 quantized with oQ (oMLX v0.5.3) at level 3.0, 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: FAILS. On the OCR fidelity check this build shifts config-table values by one column (every label carries the next column's value) while emitting well-formed output. Prefer level 5 or higher when table structure matters.

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