No official torrent yet — requesting it tells the flock what to host next. You cannot seed a catalog-only page.
Cataloged from tencent/WeVisDoc-2B
All models by tencent →English [blocked] | 简体中文 [blocked]
WeVisDoc is an end-to-end document parser for page images. Fine-tuned from Qwen3-VL-2B-Instruct and Qwen3-VL-4B-Instruct, it turns a page into structured Markdown, with LaTeX formulas and HTML tables.
WeVisDoc-4B achieves an Overall score of 95.38 on OmniDocBench v1.6 and a mean Overall score of 75.54 across the three PureDocBench tracks, ranking first among the compared end-to-end parsers in all four settings.
[Image blocked: WeVisDoc-4B leads the compared end-to-end parsers across all four reported settings.]
WeVisDoc-4B leads the compared end-to-end parsers across all four reported settings. Bars show scores on OmniDocBench v1.6 and PureDocBench Clean, Digital, and Real.
The following tables include end-to-end document parsing specialists only. WeVisDoc results are means over three inference runs.
Avg₃ is the mean of the three PureDocBench track-level Overall scores. * marks baseline results obtained with our evaluation pipeline; unmarked baseline results are taken from the corresponding papers.
Python 3.10+ is required. Install the vLLM and client dependencies:
Start the service in the first terminal:
Use Tencent/WeVisDoc-4B instead to run the 4B version.
Then process all bundled images from a second terminal:
Predictions are written to outputs/predictions/.
Replace the model ID with Tencent/WeVisDoc-4B to serve the 4B version. The launcher requires vLLM >=0.11.1. Extra arguments are passed to vLLM. To expose the service on the network, set HOST=0.0.0.0. For two GPUs and a larger context:
Use a separate virtual environment from Transformers to avoid conflicting PyTorch packages.
A client machine only needs python -m pip install -r requirements.txt. The examples expect PNG, JPEG, or WebP page images:
--image-dir processes images in that directory (not recursively) and writes one Markdown file per image, such as results/page.png.md. Existing nonempty results are skipped unless --overwrite is set.
To process every bundled image in demos/inputs/:
The demo writes one Markdown file per image to outputs/predictions/.
The client reads OPENAI_BASE_URL (default http://127.0.0.1:8000/v1), OPENAI_API_KEY (default EMPTY), and SERVED_MODEL_NAME. Match OPENAI_API_KEY to VLLM_API_KEY when authentication is enabled.
Override defaults with --base-url, --model, --timeout (600 seconds), --temperature (0), or --max-tokens (8192). Increase the token or context budget if output is truncated; reduce image size, context, or concurrency if GPU memory is insufficient.
Use a separate environment from vLLM:
Use Tencent/WeVisDoc-4B for the 4B version. --model can be omitted when WEVISDOC_MODEL_PATH is set. Local inference supports --device-map (default auto) and --max-tokens (8192). Render PDFs to page images first.
python -m venv .venvsource .venv/bin/activatepython -m pip install -r requirements-vllm.txtbash scripts/serve_vllm.sh Tencent/WeVisDoc-2Bsource .venv/bin/activatebash scripts/run_demo.shpython -m venv .venvsource .venv/bin/activatepython -m pip install -r requirements-vllm.txtbash scripts/serve_vllm.sh Tencent/WeVisDoc-2BCUDA_VISIBLE_DEVICES=0,1 TENSOR_PARALLEL_SIZE=2 MAX_MODEL_LEN=65536 \ bash scripts/serve_vllm.sh Tencent/WeVisDoc-4B --dtype bfloat16curl --fail http://127.0.0.1:8000/healthpython -m wevisdoc.client --image page.png --output results/page.mdpython -m wevisdoc.client --image-dir images --result-dir results --workers 4OPENAI_BASE_URL=http://127.0.0.1:8000/v1 \ bash scripts/run_demo.sh --workers 4python -m pip install -r requirements-local.txtpython -m wevisdoc.local --model Tencent/WeVisDoc-2B \ --image page.png --output results/page.md@article{wevisdoc, title = {WeVisDoc: From Coverage to Capability for Robust End-to-End Document Parsing}, author = {Hao Yu and Kang Liu and Linnan Zhao and Jiabo Zhan and Chong Sun and Chen Li and Jing Lyu}, year = {2026}, journal = {arXiv preprint arXiv: 2609.20423}}