nativ-community/MiniCPM-OCR-Jev-5bit

MiniCPM-OCR-Jev — MLX 5bit

Quantized from the validated BF16 conversion. Configuration: affine, 5 bits, group size 64, following nativ-community/MiniCPM-V-4.7-1B-5bit. Vision, normalization and the two-output classification score head remain unquantized. The bit label describes the quantized layers, not the average precision of the complete model.

Requires an MLX-VLM build containing the MiniCPM-OCR / MiniCPM-OCR Jev port. The runtime implementation is not bundled in this repository.

python -m mlx_vlm.models.minicpm_ocr_jev \
  --model nativ-community/MiniCPM-OCR-Jev-5bit \
  --orig-image original-crop.png --render-image rendered-ocr.png \
  --ocr-text-file candidate.txt

This model scores OCR consistency; it does not generate text. prob_1 means P(consistent). Rendering the candidate text is an upstream step.

Validation: reloaded the checkpoint and evaluated all 12 supplied image-pair/text cases. All logits and probabilities were finite. Maximum probability difference from BF16 was 0.025044; mean difference was 0.004357. At 0.5, 12/12 decisions match BF16; this is a diagnostic comparison, not a calibrated routing threshold. Quantization changes confidence scores: validate routing thresholds with the deployed precision and inputs. Per-case scores are in validation.json.

All 431 tensors outside the quantized weights match the source exactly. No performance claim is transferred from BF16 to this quantized checkpoint.

The original checkpoint was supplied for conversion; consult its original model terms. conversion.json records provenance and quantization settings, and SHA256SUMS contains file hashes.