nativ-community/MiniCPM-OCR-1.0-6bit

MiniCPM-OCR-1.0 — MLX 6bit

Quantized from the validated BF16 conversion. Configuration: affine, 6 bits, group size 64, following nativ-community/MiniCPM-V-4.7-1B-6bit. Vision, visual mergers, token embeddings, recurrent linear-attention layers and normalization 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.

mlx_vlm.generate --model nativ-community/MiniCPM-OCR-1.0-6bit \
  --image page.png --prompt "Convert this page to markdown." \
  --temperature 0 --max-tokens 4096

Validation: reloaded the checkpoint and generated outputs for three image inputs (invoice, synthetic dense page, and supplied document crop). All reached EOS. Invoice output matches BF16 after whitespace normalization: true. Exact BF16 matches across these inputs: 3/3. Quantized outputs can differ in text and formatting. These are smoke comparisons, not a comprehensive OCR accuracy benchmark. Per-input output comparisons are in validation.json.

All 683 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.