MiniCPM-OCR 1.0 DSpark

MiniCPM-OCR 1.0 DSpark

Speeds up MiniCPM-OCR 1.0 by drafting tokens that the main model checks in one pass. A draft head, not a standalone OCR model, for people who self-host document parsing.

Made by openbmb on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: MiniCPM-OCR 1.0 DSpark is a small draft head for speculative decoding, made by openbmb for the MiniCPM-OCR 1.0 document reader. It proposes several tokens at once and the main model verifies them in a single pass. It cannot do OCR alone, so you load it alongside the main MiniCPM-OCR 1.0 model.

Who it suits: It suits people who already run MiniCPM-OCR 1.0 and want documents to finish sooner. Researchers who study draft models can read its design too.

Industry | Who uses it | How they use it | Example
Document processing | Teams that parse many documents | Pair the head with MiniCPM-OCR 1.0 to finish documents sooner | Turning a batch of scanned pages into Markdown or JSON
Inference engineering | Engineers who self-host the model | Add the draft head to a serving setup for the main checkpoint | A self-hosted MiniCPM-OCR 1.0 service with this head added
Research | Researchers on speculative decoding | Study how a small draft head is built | Reading its five draft layers and 15 draft tokens per pass

You can skip it if you want an OCR model that runs alone, since this head cannot.

Where to use it:

• A self-hosted MiniCPM-OCR 1.0 service, with vLLM, SGLang, llama.cpp or Ollama
• Your own computer, on Windows, macOS or Linux with Python, to download the files
• The minicpm-ocr Python SDK, for parsing whole documents
• The project's free online playground and API for the main model, and an iOS app that runs on the device

Use case | Who | What they get back
Faster document parsing | Teams with many pages | The same parsing job with less decoding time
Self-hosted OCR service | Engineers | A serving stack where the main model verifies drafted tokens
Draft model study | Researchers | A checkpoint with documented layers and vocabulary sizes

How to set it up:

• Python and pip
• The matching main checkpoint, MiniCPM-OCR 1.0, because the head cannot run without it
• PaddlePaddle or PaddleX if you use the SDK's layout stage
• About 594 MB of disk for this head in BF16

1. Install the Hugging Face hub package.
2. Download the draft head files to your computer.
3. Get the main MiniCPM-OCR 1.0 checkpoint too.
4. Follow the project's deployment guide to serve both together.

pip install -U huggingface_hub
hf download openbmb/MiniCPM-OCR-1.0-DSpark

How to use it: You do not call the head yourself. The minicpm-ocr SDK parses a document in two stages, and the head speeds up the second one when the service is set up with it.

Tool | What it does
minicpm-ocr SDK | Runs the two stage document parsing and builds the final Markdown or JSON
PP-DocLayoutV3 | Finds text, tables, formulas and other regions, and sets reading order
MiniCPM-OCR 1.0 service | Reads each cropped region and returns its content
DSpark head | Drafts tokens that the main model verifies in one pass

One detail matters if you load the head by hand. The d2t tensor stores differences, so convert it to absolute token IDs before use:

d2t[i] + i

The file local-to-target.json holds the correct absolute IDs to check against.

Example: You send a scanned report page to the SDK. It cuts the page into regions, the service reads each region with a prompt that fits its type, and you get Markdown or JSON back in reading order.

For newcomers and experts: This is a helper, not a model you talk to. It guesses the next several tokens so the main model can check them together instead of producing them one at a time. For a first try, use the project's playground for the main model, then move to self-hosting when you need speed.

Underneath, the head has 5 layers and proposes 15 tokens per forward pass, reading hidden states from five layers of the target model. Its output vocabulary is cropped to 114,688 tokens from an input vocabulary of 248,100, and the d2t mapping links the two. The card reports a 55% gain in end to end processing efficiency with the head attached.

A page goes to the SDK
The layout model finds regions and reading order
The draft head proposes tokens
The main model verifies them in one pass
Markdown or JSON comes out

Good to know: These are the head's published specifications and speed figures.

Property | Value
Target model | MiniCPM-OCR 1.0 (1.3B)
Draft parameters | 296,837,889
Draft layers | 5
Draft tokens per forward pass | 15
Checkpoint precision | BF16

Title: Throughput with the head, as a multiple of each other system
PaddleOCR-VL-1.6: 1.49 times
MinerU2.5 Pro: 3.62 times

• License: Apache 2.0
• Cost to run: the guide offers a free online playground and API for the main model, and self-hosting needs your own hardware
• Data it reaches: downloads files from Hugging Face, and the SDK sends cropped regions over HTTP to the inference service, which can run on another machine
• Not for: standalone OCR, since the card calls it a draft model that works only alongside the main model

On Hugging Face, 13 November 2026

• Task: image text to text
• Library: transformers
• Parameters: 297M
• Weights: about 594 MB in BF16; about 148 MB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 12,549
• Likes: 14
• License: apache-2.0
• Languages: en, zh
• Based on: openbmb/MiniCPM-OCR-1.0
• Tags: transformers, safetensors, qwen3, minicpm-ocr, dspark, speculative-decoding, draft-model, multimodal, image-text-to-text

Links

• Model: https://huggingface.co/openbmb/MiniCPM-OCR-1.0-DSpark
• Files: https://huggingface.co/openbmb/MiniCPM-OCR-1.0-DSpark/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download openbmb/MiniCPM-OCR-1.0-DSpark

Then follow the usage in its model card. Nothing in this kit runs by itself.