Apache-2.0visionRuns locally4.83B (4B card safetensors)
WeVisOne-VL: Specs, License & How to Run
Tencent Apache-2.0 Qwen3-VL perception family (2B + 4B) — detect, OCR, matte.
- License
- Apache-2.0
- Parameters
- 4.83B (4B card safetensors)
- HF likes
- 91 (2026-11-01)
- First seen
- 2026-10-22
TL;DR. WeVisOne-VL is a vision model from Tencent WeChat Vision. Parameters: 4.83B (4B card safetensors).License: Apache-2.0. It is listed under remove.bg.
| Developer | Tencent WeChat Vision |
|---|---|
| Type | vision |
| Architecture | Qwen3VLForConditionalGeneration (HF model_type qwen3_vl) plus a 9,609,764-param mask_decoder.pt (Visual Feature Adapter, 6-block Box-Guided Query Decoder, FullCanvasAlphaCascadeHead). 4B text_config: 36 layers, hidden 2560, 32 heads, vocab 152680, max_position_embeddings 262144. 2B text_config: 28 layers, hidden 2048, 16 heads, same vocab and max position. Dedicated coordinate tokens <0>…<1000>; instance token <OBJ> id 152678. Not a text-to-image model. |
| Parameters | 4.83B (4B card safetensors) |
| Modalities | image, text |
| License | Apache-2.0 — LICENSE on tencent/WeVisOne-VL-4B and WeVisOne-VL-2B is Apache-2.0 with a Tencent 2026 preamble (cardData.license apache-2.0, opened 2026-11-01). GitHub Tencent/WeVisOne-VL license API is NOASSERTION; we copy the LICENSE files. Third-party components keep their original terms. Not legal advice. |
| Hardware | Not yet verified — 4B safetensors.parameters 4,830,581,248; usedStorage 29,033,805,782 bytes (~27.0 GiB). 2B safetensors.parameters 2,441,744,384; usedStorage 19,583,931,990 bytes (~18.2 GiB). Peak VRAM is not independently measured. Card: single-turn image perception; video grids are rejected by the pixel decoder. (not independently tested) |
| Hugging Face likes | 91 (as of 2026-11-01) |
| Model card | https://huggingface.co/tencent/WeVisOne-VL-4B |
| First seen | 2026-10-22 |
WeVisOne-VL (WeChat Vision All-in-One) is Tencent WeChat Vision’s unified perception family: object-level tasks (open-vocabulary detection, referring localization, counting, pointing, OCR) and pixel-level tasks (referring segmentation, promptable alpha matting) on a Qwen3-VL backbone plus a 9.6M pixel decoder. Official hosts we opened on 2026-11-01: the flagship card tencent/WeVisOne-VL-4B (created 2026-10-22T09:48:14Z; last modified 2026-10-27T04:23:41Z; 91 likes / 8,113 downloads), the 2B sibling tencent/WeVisOne-VL-2B (created 2026-10-22T09:47:50Z; last modified 2026-10-28T13:18:05Z; 67 likes / 8,817 downloads), both LICENSE files, github.com/Tencent/WeVisOne-VL (484 stars / 45 forks), the project page tencent.github.io/WeVisOne-VL, and the paper card arXiv 2610.31329. API gated was false on both cards.
This identification page hangs off remove.bg alternatives because the cards name promptable alpha matting and referring segmentation. That mapping is editorial. WeVisOne-VL is not a hosted cutout API. We have no sourced matting bake-off against remove.bg or SAM 3.
It is not Iris-3B (text-to-image). Generation compare: Iris-3B vs Qwen-Image-2.1-Turbo. A chat VLM already here is MiniCPM-V 4.7. OCR-only cards stay on Unlimited-OCR-2 and MiniCPM-OCR 1.0.
Family snapshot (2026-11-01)
| Card | Params (safetensors) | usedStorage | Likes / downloads | Created |
|---|---|---|---|---|
| WeVisOne-VL-4B | 4,830,581,248 (~4.83B) | 29,033,805,782 bytes | 91 / 8,113 | 2026-10-22 |
| WeVisOne-VL-2B | 2,441,744,384 (~2.44B) | 19,583,931,990 bytes | 67 / 8,817 | 2026-10-22 |
License on both cards: apache-2.0. Pipeline tag: image-text-to-text. Architecture class: qwen3_vl. Pixel decoder: 9,609,764 parameters (mask_decoder.pt) on both READMEs.
What we verified (2026-11-01)
| Field | 4B | 2B | Source |
|---|---|---|---|
| Host (not gated) | tencent/WeVisOne-VL-4B | tencent/WeVisOne-VL-2B | HF API |
| Last modified | 2026-10-27T04:23:41Z | 2026-10-28T13:18:05Z | same |
| LICENSE file | Apache-2.0 + Tencent preamble | same grant | raw LICENSE |
| text_config layers / hidden | 36 / 2560 | 28 / 2048 | config.json |
| max_position_embeddings | 262144 | 262144 | same |
| GitHub stars / forks | 484 / 45 | — | GitHub API |
Card intended use (copied): research in visual grounding, referring localization, pointing, counting, OCR localization, referring segmentation, and image matting. Cards say the release is not intended to make autonomous high-impact decisions or to identify individuals. Fine alpha boundaries stay sensitive to resolution, occlusion, transparency, and prompt phrasing (card limitation).
How to run (from the official README, not tested)
Install the inference package from Tencent/WeVisOne-VL, then the card’s WeVisOneVL("tencent/WeVisOne-VL-4B", device="auto", dtype="bfloat16") (or the 2B id) and model.predict(image, task="matting", query="person", temperature=0.0). forkalt has not run that snippet.
License
Both Hugging Face LICENSE files are Apache-2.0 with a Tencent 2026 preamble and a third-party-component note. GitHub’s license API is NOASSERTION; we still copy the files. Not legal advice.
What we did not invent
No quality ranking against SAM 3, remove.bg, MiniCPM-V 4.7, or Iris-3B. Peak VRAM is not yet verified. Video perception is out of scope on the cards (pixel decoder rejects video grids).
Related: SAM 3 on remove.bg alternatives — MiniCPM-V 4.7 — Mage-VL-Cowork (Microsoft CUA VLM, different job) — Unlimited-OCR-2 — PhotoCraft vs GIMP if the job is a pixel editor rather than a mask model. Catalog: home. Corrections: contact.
Deploy options
- local: Hugging Face weights (4B)
- local: Hugging Face weights (2B)
- local: GitHub inference package
- hosted: Project page
FAQ
Which checkpoint should I download?
Family page. Flagship card tencent/WeVisOne-VL-4B (created 2026-10-22T09:48:14Z; 91 likes / 8,113 downloads, 2026-11-01) is the 4.83B BF16 card. The 2B sibling tencent/WeVisOne-VL-2B (created 2026-10-22T09:47:50Z; 67 likes / 8,817 downloads) is 2.44B BF16. Same Apache-2.0 grant, same wevisone_vl predict() API, different size. Both ungated.
Is this a remove.bg clone or an image generator?
Neither as a product. The cards list open-vocabulary detection, referring localization, counting, pointing, OCR, referring segmentation, and promptable alpha matting. That is perception, not a hosted cutout site and not Iris-3B-class text-to-image. See remove.bg alternatives and Iris-3B vs Qwen-Image-2.1-Turbo.
Are the weights Apache-2.0?
Card tags apache-2.0. LICENSE files on both cards are Apache-2.0 with a Tencent preamble (Copyright 2026 Tencent) and a third-party-component note. GitHub license API is NOASSERTION. Not legal advice.
Does it match SAM 3 or remove.bg quality?
No sourced bake-off. Intended-use and limitation sentences on the cards stay upstream.
Sources & verification
- https://huggingface.co/tencent/WeVisOne-VL-4B
- https://huggingface.co/tencent/WeVisOne-VL-2B
- https://huggingface.co/tencent/WeVisOne-VL-4B/raw/main/LICENSE
- https://huggingface.co/tencent/WeVisOne-VL-2B/raw/main/LICENSE
- https://huggingface.co/tencent/WeVisOne-VL-4B/raw/main/config.json
- https://github.com/Tencent/WeVisOne-VL
- https://tencent.github.io/WeVisOne-VL
- https://huggingface.co/papers/2610.31329
License notes are not legal advice. Read the original LICENSE. Trademark & disclaimer.