# WeVisOne-VL-4B (slim)

> WeVisOne-VL-4B is an open-weight perception model of 4.8B parameters from Tencent's WeChat Vision team, built on Qwen3-VL. Name something in an image and it returns boxes, points, counts, text, masks or alpha mattes. Apache 2.0.

- Full: https://www.anchorterminal.com/tools/wevisone-vl-4b.md (~9,500 tokens) · this version ~1,780 tokens · JSON https://www.anchorterminal.com/tools/wevisone-vl-4b.json · canonical https://www.anchorterminal.com/tools/wevisone-vl-4b
- Index: https://www.anchorterminal.com/llms.txt · API: https://www.anchorterminal.com/api/v1/index.json · MCP: https://www.anchorterminal.com/mcp · Updated: 2026-11-10 · pricing https://www.anchorterminal.com/tools/wevisone-vl-4b#pricing · reviews https://www.anchorterminal.com/tools/wevisone-vl-4b#reviews
- Licence: CC BY 4.0, credit Anchor Terminal and link this page (https://www.anchorterminal.com/terms/#data)

**C · 61/100 · rank #84 of 160 open-weight models · #66 in Open-weight models · not agent-ready · confidence low**


Assessment: A small Apache 2.0 perception model that finds, counts, reads and cuts out what you name in an image, and says None when it isn't there. Boxes and points run in vLLM, SGLang or Transformers. Masks and mattes need Tencent's package and a pickle file. Two weeks old, with no questions answered yet and no safety results.

## Facts

- Kind: Open-weight model · vendor: Tencent (WeChat Vision) · category: Open-weight models · legal entity: Tencent (named in the LICENSE file's copyright line) · provenance 68/100
- Packages: huggingface `tencent/WeVisOne-VL-4B`
- Auth: None · pricing: Free · x402: no · licence: Apache-2.0
- Benchmark probes: not run yet, so Performance is pending
- Surface graded: The open weights at huggingface.co/tencent/WeVisOne-VL-4B, read from the model card, the files and their metadata, the config, the chat template, the commit history, the README of Tencent's GitHub repository for the model and the runtimes' own code at their latest releases on 10 November 2026. Tencent's company facts come from our other Tencent listings. The paper, the cookbooks and the package's code weren't read. Nothing was downloaded or run
- Model: 4,830,581,248 parameters in the safetensors file, plus a 9,609,764-parameter pixel decoder in mask_decoder.pt. Qwen3VLForConditionalGeneration with a 36-layer language model, built on Qwen3-VL by WeChat Vision at Tencent
- Files: One safetensors file in BF16 (9.66 GB), mask_decoder.pt (0.04 GB, a PyTorch pickle), chat_template.jinja, config, processor and tokenizer files and the LICENSE. No Python. Hugging Face lists a SHA-256 for the large files
- Licence: Apache 2.0 in the card's metadata, with the LICENSE file in the repository. The card's licence section speaks of Tencent's code and says third-party components keep their own licences
- Runtimes: Tencent documents its own Python package with a Transformers backend and a vLLM backend. Released Transformers 5.22.2, vLLM 0.33.0 and SGLang 0.5.23 carry the Qwen3-VL architecture, which gives the text output: boxes, points, counts and OCR. Masks and mattes need the pixel decoder, so Tencent's package
- Context: Not stated by Tencent. config.json carries Qwen3-VL's 262,144 positions and the tokenizer config says 8,192. It reads one image a turn, at up to 1,960 visual tokens in the README's results
- Hardware: No memory figure on the card. The files are 9.7 GB in BF16. The package takes a device of auto, cuda, npu or cpu
- Training data: The README describes a data engine that finds concepts from a knowledge graph, localises them with expert models checked by a vision-language model, and extends boxes to masks and mattes, with over 100 million images annotated. No sources named and no cut-off
- Tasks: Open-vocabulary detection, referring localisation, pointing, counting by box or by point, OCR and finding given text, referring segmentation and promptable alpha matting. One image and one turn at a time
- Languages: Not stated on the card or in the README
- Input and output: One image with a task and a query (a list of categories, a referring expression, or text to find). The model writes boxes and points as coordinate tokens from 0 to 1000, None when the target is absent, and an OBJ token after each box that the decoder turns into a binary mask or an alpha matte
- Recommended settings: generation_config.json sets temperature 0.7, top_p 0.8 and top_k 20. The README's results use greedy decoding with at most 1,960 visual tokens an image, and min_pixels and max_pixels are settable

- Scores: Capability pending, Reach 70, Risk 55, Agent interface 70, Documentation and upkeep 48 · total over the 7 assessed categories
- Why:  · Reach, Memory 30, the only build from Tencent is the BF16 safetensors file with its decoder, 9.7 GB, and 9.7 GB × 1.2 = 11.6 GB, within 16 GB. · Risk, File safety 15, the backbone is safetensors, but mask_decoder.pt, a PyTorch pickle the card calls a required component, sits beside it. · Agent interface, Chat template 20, chat_template.jinja in the repository (5,292 characters), Qwen3-VL's template. · Documentation and upkeep, Model card 8, Intended use, limitations and the architecture are on the card.
- Sources: 12, open questions: 6, both in the full twin
- Capabilities: model.open-weights, docs.ocr
- JSON: https://www.anchorterminal.com/api/v1/tools/wevisone-vl-4b.json
- Its vendor verifies the listing with `verify_listing` or `POST https://www.anchorterminal.com/api/v1/verify` (the full twin says how); no effect on the grade.
- Support: https://huggingface.co/tencent/WeVisOne-VL-4B/discussions; https://github.com/Tencent-Hunyuan/Hy-Memory/issues; more in the full twin

## Before you call it

1. Use Tencent's package: `WeVisOneVL("tencent/WeVisOne-VL-4B", backend="vllm")` then `model.predict(image, task=..., query=...)`
2. Tasks are detection, referring, point, count_box, count_point, ocr, ocr_grounding, segmentation and matting
3. A None answer means the target isn't in the image. Treat it as a result, not an error
4. mask_decoder.pt is a pickle. Load it only from the tencent repository, and check its SHA-256 against the Hub's
5. Send one image a turn. Video grids are rejected by the decoder
6. The card says it isn't for identifying individuals or for high-impact decisions on its own

## Connect

```bash
git clone https://github.com/Tencent/WeVisOne-VL.git && cd WeVisOne-VL
pip install -r requirements.txt && pip install -e .
# then, in Python: WeVisOneVL("tencent/WeVisOne-VL-4B", backend="transformers") or backend="vllm"
```

```bash
# No HTTP API of its own. Tencent's package calls model.predict(image, task="detection", query=[...]).
# For boxes and points only, vLLM serves the backbone at http://localhost:8000/v1 with: vllm serve tencent/WeVisOne-VL-4B
```

## Similar tools

| Tool | Grade | Score | Shared capabilities | Slim |
| --- | --- | --- | --- | --- |
| Nemotron 3.5 Parse | B | 69 | docs.ocr, model.open-weights | https://www.anchorterminal.com/tools/nemotron-3-5-parse.min.md |
| LightOnOCR-3-4B | C | 61 | docs.ocr, model.open-weights | https://www.anchorterminal.com/tools/lightonocr-3-4b.min.md |
| Surya OCR 2.5 | C | 61 | docs.ocr, model.open-weights | https://www.anchorterminal.com/tools/surya-ocr-2-5.min.md |
| WeVisOne-VL-2B | C | 61 | model.open-weights, docs.ocr | https://www.anchorterminal.com/tools/wevisone-vl-2b.min.md |
| Xiaomi-OCR-0 | C | 59 | docs.ocr, model.open-weights | https://www.anchorterminal.com/tools/xiaomi-ocr-0.min.md |


## Panel reviews (0, desk reviews from public material, no calls made)