providus-vl-mini

At a glance: 78.96 on ParseBench (mean of 4 full runs). On ParseBench's leaderboard.csv (b9ba246, 2026-10-12) that would place first among open-weight models.

Zero per-page fees: providus-vl-mini is free to use under Apache-2.0. Your only cost is your own hardware.

GGUF build: ProvidusAI/providus-vl-mini-GGUF has this model as one GGUF file (F16) for llama.cpp, LM Studio and Ollama. It needs no router and no second model. Its ParseBench score is on that page.

providus-vl-mini is a fine-tune of MinerU2.5-Pro by OpenDataLab (Apache-2.0). We did not pretrain a model. All the base knowledge comes from OpenDataLab and the MinerU team.

This repo holds the full model: MinerU2.5-Pro-2605-1.2B with our text LoRA adapter merged in (bf16, about 2.3 GB), at the repo root. The 35 MB adapter alone is in adapter/. The weights are a modified version of MinerU2.5-Pro (Apache-2.0, Section 4(b)): we merged our adapter into them. See NOTICE.

The scores below are for a setup with two models, not for this model alone: text crops go to this model, and layout, table, formula and chart requests go to the base model from OpenDataLab (see "How to use").

Base model

• Base: opendatalab/MinerU2.5-Pro-2605-1.2B, revision bff20d4ae2bf202df9f45284b4d43681555a97ed (Apache-2.0).
• Papers: MinerU2.5-Pro (arXiv:2604.04771) and MinerU2.5 (arXiv:2509.22186).

Training

A LoRA adapter fine-tuned on licensed and public-domain documents, merged into MinerU2.5-Pro-2605-1.2B. In v1, only text crops are sent to this model. Send layout, table, formula and chart requests to the base model.

Files

How to use

1. Download the base model from OpenDataLab at the revision above. The v1 setup needs it: layout, table, formula and chart requests go to it.
2. Download this repo (hf download ProvidusAI/providus-vl-mini --local-dir ./providus-vl-mini). The merged model is at the repo root. There is no merge step.
3. Serve the base model and this model with vLLM, each on its own port. Put the router in front. Text crops (Text Recognition prompts) go to this model. All other requests go to the base model. The exact commands are in serve/README.md.
4. Parse pages with providus-parse-tools (pip install providus-parse-tools==1.0.0, import providus_parse). providus-parse-tools extends mineru-vl-utils by OpenDataLab. Point providus_parse.server at the router (port 8000). Code: GitHub. Package: PyPI.

Advanced: the adapter in adapter/ is the same LoRA that is merged into the root files. You can load it on the base model with PEFT, or merge it yourself (optional step 2 of serve/README.md).

Results

Scores are the mean of 4 full ParseBench runs (± standard deviation). Scorer: ParseBench main b9ba246 (2026-10-12). Each run has 2,078 pages. No rule in our pipeline contains ParseBench documents, file names or answers. We used benchmark results to find common error types, then wrote general rules and checked them on our own data.

These scores are for the whole v1 system, not for the adapter alone: the base model, the text LoRA adapter on the text route, and providus-parse-tools 1.0.0, which also reads the PDF text layer for formatting marks (bold, italic, headings). A different pipeline will give different scores.

The system has two models. Text crops go to this model. Layout, table, formula and chart requests go to the base model (OpenDataLab, revision bff20d4ae2bf202df9f45284b4d43681555a97ed), through the router. The scores do not describe this model used alone for every kind of request.

On ParseBench's leaderboard.csv (commit b9ba246, 2026-10-12), the score of providus-vl-mini (78.96) would place 7th overall and first among open-weight models. The 7th place is 0.04 above the next row (Anthropic Fable 5.1, 78.92), which is within our run spread. providus-vl-mini is not listed in that file yet. The next open-weight row, rakedoc-nano (77.23), is licensed AGPL-3.0. providus-vl-mini is Apache-2.0, so companies can use it in closed-source products and services.

Speed by GPU

Measured on 200 ParseBench pages (page 1 of each document), one rented GPU, 2026-10-13. Batch: 24 pages at a time. One user: 50 other pages, one at a time.

Smaller GPUs are not measured yet.

A full ParseBench run (2,078 pages, multi-page documents) took 44.4 minutes from start to scores (inference and scoring, servers already running) on one H100 80 GB, 24 pages at a time. Over a full run that is about 1.28 s per page. Zero per-page fees: providus-vl-mini is free to use under Apache-2.0. Your only cost is your own hardware.

Limits

• v1 sends only text crops to this model. Tables, formulas, charts and layout come from the base model. Do not use this model alone for those requests: the scores do not cover it.
• Do not compare these scores with the leaderboard row of MinerU2.5-Pro-2605-1.2B to measure the fine-tune. That row uses MinerU's own pipeline, so the difference mixes pipeline and model effects. We have not scored the base model through providus-parse-tools yet.
• The scores are for ParseBench. Check the model on your own documents before you rely on it.
• The model can make mistakes in the text it reads. Do not use it alone for decisions where an error is costly.

Access

The files are gated on Hugging Face. Access is approved automatically when you accept the terms.

License and credits

• License: Apache-2.0. The weights are a modified version of MinerU2.5-Pro-2605-1.2B (OpenDataLab, Apache-2.0): we merged a LoRA adapter into them (Section 4(b): changes stated). Keep the NOTICE file when you redistribute the model or a derivative.
• Online services: MinerU's project licence asks online services built on MinerU to say so. If you offer an online service with this model, please state in its interface or public documentation that it uses MinerU by OpenDataLab and providus-vl-mini by Providus AI. See NOTICE.
• Thank you to OpenDataLab and the MinerU team for the base model and mineru-vl-utils.
• Thank you to the ParseBench authors, vLLM and Hugging Face PEFT.
• To cite: see CITATION.cff. Please cite the MinerU papers too.

@article{wang2026mineru25pro,
  title   = {MinerU2.5-Pro: Pushing the Limits of Data-Centric Document Parsing at Scale},
  author  = {Wang, Bin and others},
  journal = {arXiv preprint arXiv:2604.04771},
  year    = {2026}
}

@article{niu2025mineru2,
  title   = {MinerU2.5: A Decoupled Vision-Language Model for Efficient High-Resolution Document Parsing},
  author  = {Niu, Junbo and others},
  journal = {arXiv preprint arXiv:2509.22186},
  year    = {2025}
}

@article{zhang2026parsebench,
  title   = {ParseBench: A Document Parsing Benchmark for AI Agents},
  author  = {Zhang, Boyang and others},
  journal = {arXiv preprint arXiv:2604.08538},
  year    = {2026}
}