MinerU2.5-Pro-2605-0.6B

Qwen2-VL MinerU Half-ViT Half-LLM (Stage2 GKD)

A pruned, efficiency-focused variant of MinerU2.5-Pro-2605-1.2B. It keeps the exact same document-parsing protocol and checkpoint format as the official MinerU release, so existing MinerU users can switch by only changing the model path — no code or environment changes required.

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

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.

Accuracy (OmniDocBench v1.6_full)

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.

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:

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

@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},
}