Built for on-premise document intelligence: the whole system runs on one GPU, with no per-page fees.
On ParseBench's leaderboard.csv (b9ba246, 2026-10-12), providus-vl-mini would place 7th overall and first among open-weight models.
Our row is in an open pull request to ParseBench.An open-weight model with 1.2B parameters that reads the text on document pages. Run it with its base model and our open pipeline, and the system turns each page into markdown: headings stay headings, tables come out as HTML, charts come out as tables, and you get each block's position on the page. We measured it on one H100.
ParseBench overall, mean of 4 full runs
78.96± 0.04
On ParseBench's leaderboard.csv (b9ba246, 2026-10-12), providus-vl-mini would place 7th overall and first among open-weight models. providus-vl-mini is not listed in that file yet.
ParseBench overall score
Scores are the mean of 4 full ParseBench runs (± standard deviation).
The providus-vl-mini v1 system (the base model, our fine-tune on the text route, and providus-parse-tools 1.0.0) scores 78.96 ± 0.04 overall on ParseBench (scorer ParseBench main b9ba246, 2026-10-12).
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.
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.
On ParseBench's leaderboard.csv (b9ba246, 2026-10-12), providus-vl-mini would place 7th overall and first among open-weight models. providus-vl-mini is not listed in that file yet.
The leaderboard row has blank cost columns, as other open-weight rows do.
See the full leaderboardEach dot is one full ParseBench run of 2,078 pages. The four runs score from 78.91 to 79.00 overall, with a mean of 78.96 and a standard deviation of 0.04.
One NVIDIA H100 80GB HBM3, rented on RunPod (2026-10-13).
On a 200-page sample (page 1 of each document), 24 pages at a time, one rented H100 reads a page in 0.61 s. With one page at a time it takes 3.87 s per page.
That sample is 5,893 pages per hour, or 10.2 minutes per 1,000 pages.
A full ParseBench run (2,078 pages, multi-page documents, scoring included) takes 44.4 minutes from start to scores on one H100 at 24 pages at a time (inference and scoring, servers already running): 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.
Other GPUs are not measured yet. L4, RTX 4090, A100 and RTX PRO 6000 were not available on RunPod on 2026-10-13.
The system looks at the whole page and marks each block: headings, paragraphs, tables, charts and figures.
Each block is read on its own. Text comes back as text, tables as tables, and charts as tables of their values.
providus-parse-tools puts the blocks back in order and writes one markdown file. For a PDF, it also keeps bold, italic and heading marks from the text layer.
In v1, providus-vl-mini reads the text. Layout, tables, formulas and charts go to the base model from OpenDataLab, and a small router sends each request to the right one.
This is the v1 system the scores describe. It needs one GPU and vLLM. The model repo is gated: accept the terms on its page and run hf auth login first.
pip install vllm==0.28.0 pyyaml transformers==5.19.0 providus-parse-tools==1.0.0
hf download opendatalab/MinerU2.5-Pro-2605-1.2B --revision bff20d4ae2bf202df9f45284b4d43681555a97ed --local-dir ./base
hf download ProvidusAI/providus-vl-mini --local-dir ./providus-vl-mini
# start the two vLLM servers and the router: see serve/README.md, step 3
python -m providus_parse.server --port 8100 --vllm http://127.0.0.1:8000Send a page image to http://127.0.0.1:8100/predict and you get markdown back. The full guide is serve/README.md in the model repo.
GGUF builds for LM Studio, Ollama and llama.cpp are coming soon.
We are a small team at Providus. We build DocAI, and every feature in it depends on one first step: reading the page correctly. When a table comes out wrong, everything built on top of it is wrong too. We wanted a parser we could run ourselves and trust.
We did not start from zero. OpenDataLab had already released MinerU2.5-Pro as an open model, so we built on it.
providus-vl-mini is a fine-tune of MinerU2.5-Pro by OpenDataLab (Apache-2.0). We did not pretrain a model. It was fine-tuned on licensed and public-domain documents.
Around the model sits providus-parse-tools, our open pipeline that turns the model's output into markdown. providus-parse-tools extends mineru-vl-utils by OpenDataLab.
One benchmark run can get lucky. We ran the full ParseBench benchmark four times, 2,078 pages each time, with ParseBench's own scorer. We did not change the scorer.
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.
On ParseBench's leaderboard.csv (b9ba246, 2026-10-12), providus-vl-mini would place 7th overall and first among open-weight models. providus-vl-mini is not listed in that file yet.
The 7th place is 0.04 above the next row (Anthropic Fable 5.1, 78.92), which is within our run spread.
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.
You can download the weights, read the pipeline code and run the whole system on your own hardware, so your documents stay on your servers. You can also use it commercially: everything is Apache-2.0.
The whole system fits on one GPU, and there are no per-page fees. See Speed and cost for how fast it runs on one H100.
Thank you to OpenDataLab and the MinerU team for MinerU2.5-Pro and mineru-vl-utils. All the base knowledge in providus-vl-mini comes from their work. Thank you to the ParseBench authors for the benchmark, and to vLLM and Hugging Face PEFT.
Yes. The model (the base with our adapter merged in) and the adapter alone are on Hugging Face, and the pipeline is on GitHub. Access to the model files is approved automatically when you accept the terms.
Yes. The model, the adapter and the pipeline are Apache-2.0, and so is the base model. Keep the NOTICE file when you redistribute the weights.
Yes. The v1 system needs one GPU and vLLM. We measured it on one H100. Speed and cost lists the GPUs we have not measured yet.
It is the base model with our text adapter merged in. In v1 it reads the text crops; the base model still does layout, tables, formulas and charts, and our pipeline joins the results. We have not yet scored the base model through our pipeline, so we do not claim a score gain for the fine-tune alone.
The leaderboard row has blank cost columns, as other open-weight rows do. You run the model on your own hardware, so the price depends on your GPU. There are no per-page fees. Speed and cost shows how fast it runs on one H100.
Use CITATION.cff in the model repo or the pipeline repo. Please cite the MinerU2.5-Pro, MinerU2.5 and ParseBench papers too.
Apache-2.0. The weights are a modified version of MinerU2.5-Pro-2605-1.2B by OpenDataLab.
The scores are for ParseBench. Check the model on your own documents before you rely on it.