nativ-community/MiniCPM-OCR-1.0-bf16

MiniCPM-OCR 1.0 — MLX BF16

MLX conversion of the supplied MiniCPM-OCR 1.0 (1.3B) checkpoint, using a SigLIP vision encoder and Qwen3.5 language backbone. Weights are BF16, with normalization scales and recurrent parameters retained in FP32 where required. No quantization was applied.

Use an MLX-VLM build containing the minicpm_ocr model and processor port. The runtime implementation is not bundled in this checkpoint repository.

mlx_vlm.generate --model nativ-community/MiniCPM-OCR-1.0-bf16 \
  --image page.png --prompt "Convert this page to markdown." \
  --temperature 0 --max-tokens 4096

from mlx_vlm import load, generate
from mlx_vlm.prompt_utils import apply_chat_template

model, processor = load("nativ-community/MiniCPM-OCR-1.0-bf16")
prompt = apply_chat_template(
    processor, model.config, "Convert this page to markdown.", num_images=1,
)
result = generate(model, processor, prompt, image=["page.png"],
                  max_tokens=4096, temperature=0.0)
print(result.text)

The default visual compression is 4x. The compatible processor also supports 16x, image slicing and adaptive image/tile upsize. The checkpoint chat template is preserved, including its non-thinking prefix.

Validation and performance

All 779 converted tensors match the sanitized source checkpoint. The converted model transcribes the invoice fixture correctly. Sixteen preprocessing cases matched the supplied Pillow reference across both compression modes and adaptive settings. These checks are not a comprehensive OCR accuracy benchmark.

Measured on an M5 Max (40 GPU cores, 128 GB RAM), resident batch size 1, one warmup and five measured repetitions; end-to-end time includes preprocessing:

Input	Output tokens	Decode tok/s	Median time	Peak MLX memory
Synthetic invoice, 640x480, 4x	55	234.7	0.381 s	3.604 GB
Supplied document crop, 1346x230, 4x	83	225.0	0.559 s	3.680 GB
Synthetic dense page, 1440x1920, 4x	706	220.1	3.680 s	3.986 GB

On the dense-page fixture, the 4x output contained the table and omitted the heading and footer. Results depend on document layout, prompt and compression.

The source checkpoint was provided for this conversion. Consult its original model terms. File checksums are in SHA256SUMS; conversion provenance is in conversion.json.