infosave/MiniCPM-OCR-1.0-cmf

MiniCPM-OCR-1.0 — CMF

OpenBMB MiniCPM-OCR-1.0 — a 1.3B OCR model for text, tables, formulas, charts, seals and structured extraction: a SigLIP vision encoder on a Qwen3.5 language model, trained on top of MiniCPM-V 4.6. Packaged as single CMF files for cortiq, a Rust inference engine with no Python and no ML framework.

English · Русский · 中文

Quick start

cargo install cortiq-cli
hf download infosave/MiniCPM-OCR-1.0-cmf minicpm-ocr-1.0-q8_2f.cmf --local-dir .
cortiq run minicpm-ocr-1.0-q8_2f.cmf --image page.png --prompt "Text Recognition:" --max-tokens 2048

Use cortiq 0.8.54 or later; prebuilt binaries for Linux, macOS and Windows are on the releases page. --image takes a path, a data URL or an HTTP(S) URL.

Prompts

task	prompt	output
text, paragraph, title, page	Text Recognition:	plain text
table	Table Recognition:	table
formula	Formula Recognition:	LaTeX
seal	Seal Recognition:	plain text
chart	Chart to Table:	table
fields	Extract the "field", .... Format the output as a JSON object with keys: "field", .... If an item is not in the image, use "###".	JSON

The upstream SDK adds layout detection (PP-DocLayoutV3) and reading order on top of this model; these files cover the recognition step.

Files

file	language model	vision encoder	size	300 greedy tokens against the original
minicpm-ocr-1.0-q8_2f.cmf	8-bit	16-bit	1.67 GB	identical; default
minicpm-ocr-1.0-q4tp.cmf	4-bit	16-bit	1.43 GB	same text; its tokens run 48 characters short of the 300

The original is Transformers 5.21 float32 on a printed page. The preprocessing is exact and the rows the encoder hands to the language model have a cosine of at least 0.9999998.

Performance

Measured on one A4 page at 144 dpi (an overview and nine slices, 2470 image tokens):

file	stand	decode	image and prompt to the first token
q8_2f	RTX PRO 4500 Blackwell, Vulkan	254 tok/s	2.0 s
q4tp	RTX PRO 4500 Blackwell, Vulkan	267 tok/s	1.9 s
q8_2f	M4, 10-core GPU, Metal	73 tok/s	8.1 s
q4tp	M4, 10-core GPU, Metal	102 tok/s	7.7 s

The encoder takes 0.13 s per view on the card and 0.5 s on the M4; the first request on the card also uploads the weights (4 s more). On a 10-core Xeon quota with CMF_GPU=0 the language model decodes at 61 tok/s.

Server

cortiq serve minicpm-ocr-1.0-q8_2f.cmf --port 8080

curl http://localhost:8080/v1/chat/completions -H "Content-Type: application/json" -d '{
  "model": "minicpm-ocr",
  "messages": [{"role": "user", "content": [
    {"type": "image_url", "image_url": {"url": "https://example.com/page.png"}},
    {"type": "text", "text": "Text Recognition:"}
  ]}]
}'

Hardware

NVIDIA / AMD / Intel (Vulkan, DX12). The encoder and the language model run on the card; 4 GB of VRAM holds either file. Apple silicon (Metal). Both run on the GPU; a page takes about 4 GB of memory. CPU only. CMF_GPU=0.

Scope

Images only: the upstream video path, the 16x downsampling mode and the DSpark draft are not part of these files.

License

Apache-2.0, as the original.

---

Документация на русском

MiniCPM-OCR-1.0 от OpenBMB — OCR-модель на 1.3B для текста, таблиц, формул, диаграмм, печатей и извлечения полей: энкодер SigLIP и языковая модель Qwen3.5. Файлы CMF для cortiq, движка на Rust без Python.

cargo install cortiq-cli
hf download infosave/MiniCPM-OCR-1.0-cmf minicpm-ocr-1.0-q8_2f.cmf --local-dir .
cortiq run minicpm-ocr-1.0-q8_2f.cmf --image page.png --prompt "Text Recognition:" --max-tokens 2048

Нужен cortiq 0.8.54 или новее. Подсказки: Text Recognition:, Table Recognition:, Formula Recognition:, Seal Recognition:, Chart to Table:. Детектор разметки и порядок чтения из SDK вендора в файлы не входят: здесь шаг распознавания.

q8_2f на странице A4 выдаёт те же 300 токенов, что и оригинал в float32; q4tp меньше и даёт тот же текст на этой странице. Скорость на RTX PRO 4500: 254 и 267 ток/с, на M4 с 10-ядерным GPU (Metal): 73 и 102 ток/с. Только CPU: CMF_GPU=0. Лицензия Apache-2.0.

---

中文说明

MiniCPM-OCR-1.0 是 OpenBMB 的 1.3B OCR 模型，用于文本、表格、公式、图表、印章和字段提取：SigLIP 视觉编码器加 Qwen3.5 语言模型。这里是 cortiq（无 Python 的 Rust 推理引擎）使用的 CMF 文件。

cargo install cortiq-cli
hf download infosave/MiniCPM-OCR-1.0-cmf minicpm-ocr-1.0-q8_2f.cmf --local-dir .
cortiq run minicpm-ocr-1.0-q8_2f.cmf --image page.png --prompt "Text Recognition:" --max-tokens 2048

需要 cortiq 0.8.54 及以上。提示词：Text Recognition:、Table Recognition:、 Formula Recognition:、Seal Recognition:、Chart to Table:。上游 SDK 的版面检测和 阅读顺序不包含在内，这些文件负责识别这一步。

在一页 A4 上，q8_2f 的 300 个贪心 token 与 float32 原模型一致；q4tp 更小，这一页的文本 相同。RTX PRO 4500 上的速度：254 和 267 tok/s；10 核 GPU 的 M4（Metal）：73 和 102 tok/s。 仅 CPU：CMF_GPU=0。许可证 Apache-2.0。