aijiag

MinerU2.5-Pro-2605-0.6B

README

License: apache-2.0

Why this model

~2× faster than the official 1.2B model on the same hardware, while retaining 94.82 OmniDocBench overall (vs. 95.72 official).
Built by structured pruning of Qwen2-VL (ViT 32→15 layers, LLM 24→12 layers), then trained on the same MinerU training pipeline and data, ensuring reliable, format-consistent output.
Ideal for users who prioritize throughput and accept a small accuracy trade-off.

Model summary

Table with columns: Official MinerU2.5-Pro, This model
	Official MinerU2.5-Pro	This model
Parameters	1.2B	0.64B
ViT layers	32	15
LLM layers	24	12
Checkpoint format	legacy (transformers 4.x)	legacy (transformers 4.x) — same
Tokenizer / special tokens	MinerU protocol	MinerU protocol — same
Training data	MinerU data engine	MinerU stage-2 GKD (same data)

Speed

Measured on a single NVIDIA A100-SXM4-80GB, full OmniDocBench v1.6 (1,651 original pages), mineru-vl-utils 1.0.5, identical inference config.

Table with columns: Setup, pages/s, vs. official, total tok/s
Setup	pages/s	vs. official	total tok/s
MinerU2.5-Pro official + vLLM 0.11	2.22	1.00×	14,164
This model + vLLM 0.11	2.63	1.18×	16,768
This model + vLLM 0.11 + `MINERU_INTRA_OP_NUM_THREADS=1 OMP_NUM_THREADS=1`	3.60	1.62×	23,006
This model + vLLM 0.24	4.69

All four rows run on the same legacy checkpoint — a drop-in replacement for official MinerU. The last row shows the additional headroom available when also upgrading to vLLM 0.24, which loads this legacy format directly (no checkpoint changes required).

Accuracy (OmniDocBench v1.6_full)

Table with columns: Metric, Official MinerU2.5-Pro, This model
Metric	Official MinerU2.5-Pro	This model
Overall ↑	95.72	94.82
Text Edit ↓	0.036	0.0431
Formula CDM ↑	97.15	96.82
Table TEDS ↑	93.62	91.96
Table TEDS-S ↑	96.01	95.14
Read Order Edit ↓	0.123

Official numbers are from the MinerU2.5-Pro model card; this model's numbers are measured on the same OmniDocBench v1.6 benchmark.

Quick start

Drop-in replacement for MinerU2.5-Pro. Use the same mineru-vl-utils + vLLM stack, only change the model path.

python
from vllm import LLM
from PIL import Image
from mineru_vl_utils import MinerUClient, MinerULogitsProcessor

llm = LLM(
    model="<this-model-path>",
    logits_processors=[MinerULogitsProcessor],  # if vllm>=0.10.1
)

client = MinerUClient(backend="vllm-engine", vllm_llm=llm)
print(client.two_step_extract(Image.open("/path/to/page.png")))

Requirements

vllm >= 0.10.1 (0.11.x recommended)
mineru-vl-utils >= 1.0.5 (required for best accuracy)
transformers >= 4.56

Throughput tip

To maximize throughput on vLLM 0.11.x, set:

bash
export MINERU_INTRA_OP_NUM_THREADS=1
export OMP_NUM_THREADS=1

This lifts the model from 2.63 → 3.60 pages/s (+37%) by eliminating CPU thread contention during image preprocessing.

Citation

BibTeX
@misc{wang2026mineru25propushinglimitsdatacentric,
      title={MinerU2.5-Pro: Pushing the Limits of Data-Centric Document Parsing at Scale},
      author={Bin, Wang and Tianyao, He and Linke, Ouyang and Fan, Wu and Zhiyuan, Zhao and Tao, Chu and Yuan, Qu and Zhenjiang, Jin and Weijun, Zeng and Ziyang, Miao and Bangrui, Xu and Junbo, Niu and others},
      year={2026},
      eprint={2604.04771},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2604.04771},
}

Available on FriendliAI

Dedicated Endpoints

Run this model inference on single tenant GPU with unmatched speed and reliability at scale.

Learn more

Container

Run this model inference with full control and performance in your environment.

Learn more

Model Details

Model Provider

aijiag

Model Tree

Base

this model

Input Modalities

TextImage

Output Modalities

Text

Supported Functionality

Dedicated EndpointsContainer

Explore FriendliAI today

Get started Talk to an engineer