MiniCPM-OCR-1.0-DSpark OCR Acceleration Draft Model: 4GB+ VRAM

MiniCPM-OCR-1.0-DSpark OCR Acceleration Draft Model: 4GB+ VRAM

At a Glance

Item Value
Repository openbmb/MiniCPM-OCR-1.0-DSpark
Family guide MiniCPM-OCR-1.0 guide (1 articles)
Publisher guide OpenBMB: models and licenses
Published 2026-10-26
License apache-2.0
Formats safetensors
Source type Primary source (the publisher itself)

Values determined by this site’s code when the information was collected. Dates are JST.

Overview

openbmb/MiniCPM-OCR-1.0-DSpark is a DSpark draft (speculative decoding) head designed to accelerate the decoding process of MiniCPM-OCR 1.0. What is provided in this repository is a draft model, not a standalone OCR model. It must always be used in combination with the corresponding main checkpoint.

DSpark improves decoding speed by utilizing a mechanism where a lightweight draft model proposes tokens, and then the main model verifies those proposals in a single pass. It is reported that introducing this head improves the end-to-end processing efficiency of MiniCPM-OCR 1.0 by 55%.

Specifications

Property Value
Target model MiniCPM-OCR 1.0 (1.3B)
Draft parameters 296,837,889
Draft layers 5
Draft tokens per forward pass 15
Target hidden-state layers [3, 7, 11, 15, 19]
Checkpoint precision BF16
Input vocabulary 248,100
Draft output vocabulary 114,688

This table shows that the DSpark head has a very lightweight design. The number of draft parameters is approximately 297 million, and the number of layers is extremely small at 5. In addition, it is configured to propose 15 draft tokens in a single forward pass. The input vocabulary size is 248,100, while the draft output vocabulary size is limited to 114,688.

Note that this checkpoint contains d2t offset mappings, but since they are delta-encoded (target_id - draft_id), they must be converted to absolute IDs (d2t[i] + i) before use. This conversion is already done in GGUF builds.

Performance

Regarding the performance of this repository, it is explained from two aspects: the performance of the base MiniCPM-OCR 1.0 and the acceleration effect by DSpark.

First, regarding the performance of the original model MiniCPM-OCR 1.0, the measurement results published by the creators are introduced. It recorded a score of 93.68 on Real5-OmniDocBench, a benchmark targeting real-world document analysis. It also achieved a score of 54.7 on OCRBench v2. Regarding throughput, it is reported to reach 1.2 pages/s in an environment using an NVIDIA GeForce RTX 5090.

Next is the acceleration effect by the DSpark draft head. By using this head, end-to-end processing efficiency improves by 55%. In terms of specific throughput comparison, it is said to reach about 1.49x that of PaddleOCR-VL-1.6 and about 3.62x that of MinerU2.5 Pro.

Strengths and Use Cases

This checkpoint is a draft head for accelerating inference of MiniCPM-OCR 1.0 via speculative decoding. It is not a standalone model and must always be used in combination with the corresponding main model.

The original model, MiniCPM-OCR 1.0, is a vision-language model specialized in document analysis and document understanding, primarily supporting Chinese and English documents. It excels at recognition from diverse images in real-world environments, such as scanned images, curved pages, photographed documents, and tilted pages.

Specifically, by specifying prompts according to the cropped region or use case, it supports tasks such as the following:

  • Text Recognition (Prompt: Text Recognition:): Recognition of regular sentences, paragraphs, and titles (plain text output)
  • Table Recognition (Prompt: Table Recognition:): Recognition of table structures (HTML format output via SDK post-processing)
  • Formula Recognition (Prompt: Formula Recognition:): Recognition of inline and standalone mathematical formulas (LaTeX format output)
  • Seal Recognition (Prompt: Seal Recognition:): Character recognition inside seals
  • Chart Recognition (Prompt: Chart to Table:): Data extraction from charts and conversion to table format

Furthermore, it supports not only document-wide paper recognition but also Key Information Extraction (KIE), which directly extracts specific fields (such as a store name or amount on a receipt) and outputs them in JSON format.

Hardware Requirements

Estimated requirements (calculated by Local Model Watch) — 297M parameters

Your VRAM Quantization File size Est. memory needed
4GB (laptop iGPU / phone class) I64 0.6GB 0.7GB

Inference engine support (architecture name matched against each project’s own model registry in its source code, checked 2026-11-17): llama.cpp: registered, vLLM: registered, MLX (mlx-lm): registered.

Memory estimates add a 20% runtime overhead (KV cache, etc.) to the actual size of the distributed files. Actual usage varies with context length, batch size and inference engine. These figures are computed by this site from file sizes, not published by the model’s authors. Compare with other models in our VRAM quick reference. What the quantization names mean: glossary.

Can You Run It Locally?

Runs in Ollama, LM Studio and llama.cpp via a converted build.

The publisher ships safetensors, but openbmb/MiniCPM-OCR-1.0-DSpark-gguf provides a GGUF build you can use.

License — apache-2.0 (Commercial use allowed): Permits commercial use, modification and redistribution. Redistribution requires including the license and stating changes; includes a patent grant.

Compiled by this site’s code from the published formats, converted builds we have found, and each engine’s own model registry. “Not found" means we have not seen such a build, not that none exists. License summaries are not legal advice — check the publisher’s original terms before relying on them.

Recent Models in the Same Size Class

Models with up to 4B parameters that Local Model Watch covered recently, listed by code from our article log for comparison. VRAM tiers are this site’s estimates; licenses are as stated on the model cards.

→ Scroll horizontally to see all columns

Model Parameters Smallest VRAM tier License Our article
LiquidAI/d1-omni-600M 587M 4GB lfm1.0 d1-omni-600M Multimodal Decision Model: 4GB+ VRAM, GGUF Builds (2026-11-17)
LiquidAI/d1-3B 3.1B 4GB lfm1.0 d1-3B Vision-Language Model: 4GB+ VRAM, GGUF Builds (2026-11-17)
baidu/Unlimited-OCR-2 3.3B 8GB mit Unlimited-OCR-2 Vision-Language Model: 8GB+ VRAM (2026-11-17)
openbmb/MiniCPM-OCR-Jev-GPTQ 2.2B 4GB apache-2.0 OpenBMB Releases Quantized Versions of MiniCPM-OCR-Jev (2026-11-17)
LiquidAI/d1-2B-GGUF 1.9B 4GB lfm1.0 d1-2B-GGUF Text, Image, and Video Decision-Making Model: 4GB+ VRAM (2026-11-17)

How to Get It

This model is distributed on Hugging Face in safetensors format (BF16). In addition, GGUF format draft models are also publicly available.

An example command for serving with vLLM by combining the main model (openbmb/MiniCPM-OCR-1.0) and this draft model with speculative decoding enabled is as follows:

vllm serve openbmb/MiniCPM-OCR-1.0 \
  --served-model-name MiniCPM-OCR \
  --trust-remote-code \
  --max-model-len 40960 \
  --max-num-batched-tokens 40960 \
  --max-num-seqs 1 \
  --trust-request-mm-kwargs \
  --mm-processor-kwargs '{"downsample_mode":"4x","max_slice_nums":9}' \
  --speculative-config '{"method":"dspark",
                         "model":"openbmb/MiniCPM-OCR-1.0-DSpark",
                         "num_speculative_tokens":15,
                         "enable_adaptive_verification":false}' \
  --default-chat-template-kwargs '{"enable_thinking":false}'

Across the ecosystem, deployment guides are available for various runtimes and environments such as SGLang, llama.cpp, Ollama, Transformers, and FlagOS, in addition to vLLM.

When using DSpark, make sure to operate the corresponding main model and draft model from the same release as a pair. Please note that they cannot be combined with checkpoints of different versions.

Quantized and Converted Variants

→ Scroll horizontally to see all columns

Added Publisher Format Repository Smallest VRAM tier (build, est. memory)
2026-10-27 openbmb GGUF openbmb/MiniCPM-OCR-1.0-DSpark-gguf F16 0.7GB (fits in 4GB VRAM)

File sizes of each build:

  • Available builds in openbmb/MiniCPM-OCR-1.0-DSpark-gguf: Q4_0 0.2GB / Q4_K_S 0.2GB / Q4_K_M 0.2GB / Q4_1 0.2GB / Q5_0 0.2GB / Q5_K_S 0.2GB / Q5_K_M 0.2GB / Q5_1 0.2GB / Q6_K 0.2GB / Q8_0 0.3GB / F16 0.6GB

This section is appended automatically by Local Model Watch when a converted build of this model appears after publication. Memory figures are estimated from the size of the distributed files. What the quantization names mean: glossary.

Related Articles

What to Read Next

Sources