openbmb/MiniCPM-OCR-1.0-AWQ

Demo (CN / Global) · GitHub · Free API · SDK usage · DSpark · Skill

This repository hosts the AWQ W4A16 quantized version of MiniCPM-OCR 1.0. For the original BF16 weights and the base model card, see openbmb/MiniCPM-OCR-1.0. Everything below describes the model itself; see Quantization for what changed. MiniCPM-OCR 1.0 (1.3B) is the first OCR-specific model in the MiniCPM-V/o family. It focuses on text-centric document understanding — text, tables, formulas, charts, seals, and structured information extraction — and balances quality against speed on real-world captures.

Highlights

• 🏆 SOTA document parsing: 93.68 on Real5-OmniDocBench, a benchmark for real-world document parsing, with leading results on OmniDocBench v1.6, Wild-OmniDocBench, and olmOCR-Bench. Handles Chinese and English documents and PDFs, including scans, curved pages, photographs, skewed pages, and recaptured documents.
• ⚡ Efficient inference. With DSpark decoding acceleration (BF16, GGUF) end-to-end processing efficiency improves by 55%. Throughput is about 1.49× PaddleOCR-VL-1.6 and 3.62× MinerU2.5 Pro.
• 🔍 General recognition and information extraction: 54.7 on OCRBench v2, plus key information extraction (KIE) for downstream automation.
• 🛠️ Easy to deploy and develop: vLLM, SGLang, llama.cpp, and Ollama from day one, with quantized variants and multi-chip builds. Supports further fine-tuning, and ships a pip SDK, a full API, and agent workflows. Online demos: China · Global.

SOTA quality at higher throughput

Among the open-source OCR models compared, MiniCPM-OCR 1.0 reaches 1.2 pages/s and the highest equal-weight average score across OmniDocBench v1.6, Real5-OmniDocBench, Wild-OmniDocBench, and olmOCR-Bench.

_{Average score is the equal-weight mean of the four benchmarks. All models use accelerated configurations. Settings and reproduction: reproduction guide. Per-model throughput: model guide.}

Benchmark comparisons

Full MiniCPM-OCR 1.0 evaluation results

Each panel keeps its own benchmark's scores; higher is better. Scores from different benchmarks are not directly comparable. Test settings and reproduction steps are in the reproduction guide; per-benchmark discussion is in the model guide.

Throughput

On an NVIDIA GeForce RTX 5090 with all 1651 OmniDocBench v1.6 images, image-level concurrency 1, and each system's accelerated configuration, MiniCPM-OCR 1.0 reaches 1.2 pages/s (relative throughput 1.00×; higher is faster).

Concurrency 1 is image-level: regions inside a page are still recognised in parallel. Real throughput also depends on page content, backend, concurrency, and hardware.

Examples

Left to right: original document, detected layout, and rendered Markdown.

English documents

Chinese documents

More examples

Original documents and parsed results, side by side.

English documents

Chinese documents

Model architecture

Item	Description
Parameters	1.3B
Base	Trained for OCR on top of MiniCPM-V 4.6
Components	SigLIP vision encoder + Qwen3.5 language model
Primary languages	Chinese and English
Tasks	Text, tables, formulas, charts, seals, and key information extraction
Parsing pipeline	MiniCPM-OCR 1.0 + PP-DocLayoutV3 layout model + minicpmvocr SDK
Visual tokens	4× compression, scale_resolution 448, up to 9 slices

See the model guide for architecture, parsing workflow, deployment options, and benchmark breakdowns.

Quick start

Use with the MiniCPM-OCR SDK

Use the official minicpm-ocr SDK to parse local images or PDFs with this AWQ checkpoint. The SDK detects layout with PP-DocLayoutV3, recognises each region, and saves structured JSON and Markdown in reading order.

Install Python 3.10+ and the SDK:

pip install "minicpm-ocr[all]"

Install minicpm-ocr for your platform using the installation guide.

Before running the SDK, start a vLLM recognition service for the AWQ checkpoint in this repository, openbmb/MiniCPM-OCR-1.0-AWQ, following the vLLM deployment guide. Use this repository ID as the served model path, --served-model-name minicpm-ocr, --port 8080, and --max-model-len 40960. The service reads the AWQ quantization settings from this repository's config.json.

The SDK examples use http://127.0.0.1:8080, model alias minicpm-ocr, the vllm backend, and an NVIDIA GPU for layout detection. Match the URL, alias, and context length to your service. The SDK connects to the running service; it does not load the quantized weights into the SDK process.

CLI

Save a local image as demo.png, then run:

minicpmvocr parse demo.png \
  --backend vllm \
  --server-url http://127.0.0.1:8080 \
  --model minicpm-ocr \
  --device cuda:0 \
  --set engine.max_model_len=40960 \
  --set run.run_root=output \
  --output output/demo.md

The CLI automatically downloads PP-DocLayoutV3 to models/PP-DocLayoutV3_infer/ when needed. The final Markdown is saved to output/demo.md; structured JSON and per-page Markdown are kept in output/<run_tag>/results_json/ and output/<run_tag>/output/.

Python API

The same pipeline is available from Python:

from importlib.resources import files
from pathlib import Path

from minicpmvocr import load_config, pipelined_run
from minicpmvocr.layouts.download import ensure_layout_model


def main():
    layout_model = ensure_layout_model(Path("models/PP-DocLayoutV3_infer"))

    cfg = load_config(
        str(files("minicpmvocr").joinpath("configs/minicpmvocr.yaml")),
        overrides={
            "input": {"file_path": str(Path("demo.png").resolve())},
            "engine": {
                "type": "vllm",
                "server_url": "http://127.0.0.1:8080",
                "model_name": "minicpm-ocr",
                "max_model_len": 40960,
            },
            "layout": {
                "model_dir": str(layout_model),
                "device": "cuda:0",
            },
            "pipeline": {"overlap_layout": True},
            "run": {"run_root": str(Path("output").resolve())},
        },
    )

    pipelined_run(cfg)

    print("JSON directory:", cfg["_meta"]["result_json_dir"])
    print("Markdown directory:", cfg["_meta"]["output_dir"])


if __name__ == "__main__":
    main()

ensure_layout_model() reuses or downloads the official layout model. Results are written to output/<run_tag>/results_json/ and output/<run_tag>/output/. To parse a PDF, replace demo.png with its local path.

For CPU layout detection, use --device cpu in the CLI or set layout.device to cpu in Python. This changes only the layout detector; the recognition service still uses the deployed quantized model.

Run minicpmvocr parse --help for CLI options, or see the SDK API reference for configuration and document-level Markdown export.

Prompts

The prompt selects the task and must match the region you cropped:

Region	Prompt	Output
Text block / paragraph / title / page	Text Recognition:	plain text
Table	Table Recognition:	HTML (via SDK post-processing)
Formula	Formula Recognition:	LaTeX
Inline formula	Text Recognition:	LaTeX
Seal	Seal Recognition:	seal text
Chart	Chart to Table:	table

If finish_reason is length, the output was truncated — raise the token budget.

Key information extraction (KIE)

MiniCPM-OCR 1.0 can pull specified fields straight out of an image — for example the store name and amount on a receipt. Send the image with a field-extraction prompt and ask for a JSON object; you do not need to produce page Markdown first.

Use the image recognition API example and replace Text Recognition: with:

Extract the "store_name", "amount". Format the output as a JSON object with keys: "store_name", "amount". If an item is not in the image, use "###".

Replace both field lists with your own field names (Chinese or English). The response is model-generated text: parse the JSON and check fields and values before using it downstream.

Agent skill

Have a coding agent follow the skill to set up the environment, download the model, and parse a first document:

Read and execute https://github.com/OpenBMB/MiniCPM-OCR/blob/main/docs/SKILL_en.md

Fine-tuning

LLaMA-Factory · ms-swift — data preparation, full-parameter SFT with the multimodal projector frozen, training, and evaluation (French handwritten text lines).

Quantization

Field	Value
Method	AWQ (quant_method: awq, version: gemm, zero_point: true)
Weight bits	4
Group size	128
Activations	BF16 (W4A16)
Not quantized	lm_head, vision tower, merger, in_proj_a / in_proj_b
Tool	tc-mb/AutoAWQ (AutoAWQ fork)
Calibration	234 document-style samples derived from OmniDocBench annotations

Only the language-model linear layers are quantized. The vision encoder, multimodal projector, lm_head, and the GatedDeltaNet in_proj_a / in_proj_b projections stay in BF16, so image encoding is unchanged and recurrent gating is preserved.

Weight footprint drops from about 2.42 GiB (BF16) to about 1.74 GiB.

File	Notes
model.safetensors	4-bit quantized weights, BF16 vision tower
config.json	base config + quantization_config
tokenizer.json, tokenizer_config.json, chat_template.jinja	tokenizer and chat template
preprocessor_config.json, processor_config.json	image preprocessing and processor configs
generation_config.json	generation defaults

quantization_config in config.json is what vLLM reads; no extra flag is needed.

Related checkpoints

Format	Hugging Face	ModelScope
BF16 (original)	🤗	ModelScope
GGUF (llama.cpp)	🤗	ModelScope
AWQ W4A16 (this repo)	🤗	ModelScope
GPTQ W4A16	🤗	ModelScope
DSpark draft (BF16)	🤗	ModelScope
DSpark draft (GGUF)	🤗	ModelScope

DSpark requires a matching main and draft model from the same release — do not mix draft weights with a different main checkpoint. See DSpark acceleration.

For OCR quality assessment (auto-accept, human-review routing, best-of-N across models), see MiniCPM-OCR Jev.

FAQ

Installation, model downloads, GPU memory, input formats, and unexpected outputs: FAQ.

License

Model weights, code, and documentation are released under the Apache License 2.0.