vLLM/Recipes
Search
MiniCPM (OpenBMB)

openbmb/MiniCPM-OCR-1.0

MiniCPM-OCR 1.0 (1.3B) — OCR-specific document-parsing VLM that turns page images and region crops into Markdown text, tables, formulas, charts, and seals

93.68 on Real5-OmniDocBench — OCR-specific 1.3B model with DSpark speculative decoding

dense1.3B262,144 ctxvLLM 0.34.0+Multimodal
Default configuration: 1× RTX 4090 · Tensor Parallel · BF16
vllm serve openbmb/MiniCPM-OCR-1.0 \
  --trust-request-mm-kwargs \
  --limit-mm-per-prompt '{"image":36,"video":0}' \
  --max-model-len 40960 \
  --max-num-batched-tokens 40960 \
  --tensor-parallel-size 1 \
  --speculative-config '{"method":"dspark","model":"openbmb/MiniCPM-OCR-1.0-DSpark","num_speculative_tokens":15,"enable_adaptive_verification":false}'

The interactive builder loads with JavaScript. Append .md to this page's URL, query string included, for any configuration's command as Markdown — e.g. /openbmb/MiniCPM-OCR-1.0.md.

Overview

MiniCPM-OCR 1.0 is a 1.3B OCR-specific model in the MiniCPM-V family. It pairs a SigLIP-style vision encoder with the Qwen3.5-0.8B hybrid linear/full-attention backbone from MiniCPM-V 4.6, and is trained for text-centric document parsing: text blocks, tables, formulas, charts, seals, and key-information extraction. It scores 93.68 on Real5-OmniDocBench and 54.7 on OCRBench v2.

The checkpoint is the recognition stage. Send a page image or a cropped region together with a region prompt; for PDFs, the companion minicpm-ocr SDK detects layout with PP-DocLayoutV3, crops regions, calls this service, and assembles the results into Markdown/JSON.

This recipe was run end to end on one RTX 4090 (24 GB) with vLLM 0.34.0 and transformers 5.22, on the BF16 checkpoint and both official 4-bit checkpoints, with and without the DSpark draft model. With DSpark on the 4090, test requests accepted ~60% of drafted tokens (348/570), and multi-token requests completed 20–45% faster than plain decoding.

Prerequisites

  • vLLM ≥ 0.34.0 — v0.33.0 registered MiniCPMOcrForConditionalGeneration but pins transformers < 5.21, before the minicpm_ocr processors shipped. 0.34.0 includes the follow-up downsample fix (PR #64105).
  • transformers ≥ 5.21, < 5.23 — the range the model's processors need within vLLM 0.34.0's pin.
  • One GPU is enough: BF16 weights are ~2.6 GB, and the DSpark draft model adds ~0.6 GB. Use temperature=0 for OCR.
  • Context — the default command keeps the model guide's 40,960-token window, sized for page crops; the checkpoint supports up to 262,144 tokens.

Client usage

The prompt selects what to read in the image: Text Recognition: for text blocks and pages, Table Recognition: for tables, Formula Recognition: for formulas, Seal Recognition: for seals, and Chart to Table: for charts. Key-information extraction takes a field list instead. Tables and charts come back as OTSL; the SDK converts them to HTML.

from openai import OpenAI

client = OpenAI(api_key="EMPTY", base_url="http://localhost:8000/v1", timeout=600)

response = client.chat.completions.create(
    model="openbmb/MiniCPM-OCR-1.0",
    messages=[{
        "role": "user",
        "content": [
            {"type": "image_url", "image_url": {"url": "https://.../page.png"}},
            {"type": "text", "text": "Text Recognition:"},
        ],
    }],
    temperature=0.0,
    max_tokens=2048,
)
print(response.choices[0].message.content)

Full documents with the SDK

The official SDK handles layout detection, region cropping, and result assembly around this service. Install it in a separate environment (pip install "minicpm-ocr[all]") and point it at the server. Pass the served model name (--model) to match the server; the SDK defaults to minicpm-ocr.

The server keeps --trust-request-mm-kwargs so the SDK can set per-request image-processing options such as downsample_mode. Keep it only for trusted clients: without it, vLLM rejects per-request image options.

References