providus-vl-mini-GGUF

This repo holds providus-vl-mini in GGUF format, for llama.cpp and LM Studio. The weights are the merged model of ProvidusAI/providus-vl-mini, converted to F16 (no quantization).

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. See NOTICE.

Files

How to use

One model for every request

The GGUF build is one model. In the pipeline, every request (layout, text, table, formula, chart) goes to it. There is no router and no second model. This is different from the base models setup of the main repo, where the base model from OpenDataLab answers the layout, table, formula and chart requests. The scores below were measured in the one-model setup.

llama.cpp

Download both files into one folder (hf download ProvidusAI/providus-vl-mini-GGUF --local-dir ./providus-vl-mini-GGUF), then run:

llama-server -m providus-vl-mini-F16.gguf --mmproj mmproj-providus-vl-mini-F16.gguf \
  --alias providus-vl-mini --host 127.0.0.1 --port 8000 \
  -c 8192 -np 1 --special --image-min-tokens 64 --image-max-tokens 2048 --repeat-last-n 8192

• --mmproj loads the vision projector.
• --special prints all special tokens. The layout and table tokens are marked as user-defined tokens in this file, so llama.cpp, LM Studio and Ollama print them even without this flag.
• --image-min-tokens and --image-max-tokens set the image size limits of the model. They make llama.cpp resize the crops as the model expects.
• -c 8192 -np 1 gives one request a context of 8192 tokens. For N requests at the same time, use -c 8192 times N, and -np N.
• --repeat-last-n 8192 sets the length of the repeat-penalty window to the context.
• On a CUDA GPU with many slots (the full run used -c 196608 -np 24), add --cache-ram 0 and start the server with GGML_CUDA_DISABLE_GRAPHS=1. Without them, llama-server stopped with a CUDA error in the vision encoder in our full runs.

Then parse pages with providus-parse-tools (pip install providus-parse-tools==1.0.0):

python -m providus_parse.server --port 8100 --vllm http://127.0.0.1:8000

The --vllm argument takes the address of any OpenAI-compatible server, here llama-server. Code: GitHub.

LM Studio

Keep the two GGUF files in one folder. LM Studio expects ~/.lmstudio/models/<publisher>/<model>/<file>.gguf. Download them with lms get https://huggingface.co/ProvidusAI/providus-vl-mini-GGUF, or search for ProvidusAI/providus-vl-mini-GGUF in the app. Load providus-vl-mini and start the local server (Developer tab, or lms server start). Set the context length to 8192. Then point providus_parse.server at the server. LM Studio lists every model you have, so name this one: MINERU_VL_MODEL_NAME=providus-vl-mini python -m providus_parse.server --port 8100 --vllm http://127.0.0.1:1234.

Ollama

Put Modelfile from this repo in the folder with the two GGUF files, then run ollama create providus-vl-mini -f Modelfile and MINERU_VL_MODEL_NAME=providus-vl-mini:latest python -m providus_parse.server --port 8100 --vllm http://127.0.0.1:11434.

Ollama runs this model on its own engine and resizes the images with its own code, so its output can differ from llama.cpp. The scores below do not cover Ollama. For the measured setup, use llama.cpp or LM Studio.

Results

Tested with: LM Studio (llama.cpp runtime 2.57.0), macOS: 3 pages parsed.

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.

The first row is one full run, so it has no standard deviation. The second row is the base models setup of the main repo: the MinerU2.5-Pro base model and providus-vl-mini on vLLM (mean of 4 runs, standard deviation 0.04 on the overall score).

• The GGUF run used llama.cpp b11609 (commit fcac190), build cuda-sm90, on NVIDIA H100 80GB HBM3.
• The base models row is a different setup: the MinerU2.5-Pro base model answers the layout, table, formula and chart requests, and providus-vl-mini reads the text.
• A different pipeline will give different scores.

Limits

• The scores are for ParseBench, in the one-model setup. Check the model on your own documents before you rely on it.
• The base models setup of the main repo sends only text crops to the merged model. The GGUF build answers every kind of request, and that setup was measured as shown above.
• The model can make mistakes in the text it reads. Do not use it alone for decisions where an error is costly.
• llama.cpp resizes images with its own code, so a crop can differ a little from the Hugging Face processor.
• In the vocabulary of providus-vl-mini-F16.gguf, the layout and table tokens (<|box_start|>, <|ref_start|>, <fcel>, <nl> and others) are marked as user-defined, not control, tokens. The weights and token ids are the same as in the scored file.

Access

The files are open: you need no login and no terms to download them.

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 and converted them to GGUF. Keep the NOTICE file when you redistribute the model or a derivative.
• Thank you to OpenDataLab and the MinerU team for the base model and mineru-vl-utils, to the llama.cpp authors, and to the ParseBench authors.
• To cite: see CITATION.cff. Please cite the MinerU papers too (see the model card of the main repo).