# WeVisOne-VL

> Source: https://aiwiki.ai/wiki/wevisone_vl
> Updated: 2026-11-04
> Categories: AI Models, Computer Vision, Multimodal AI, Open Source AI
> License: CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/) - attribute to "AI Wiki (aiwiki.ai)"
> Cite as: AI Wiki. "WeVisOne-VL." aiwiki.ai, 4 Nov 2026. https://aiwiki.ai/wiki/wevisone_vl
> From AI Wiki (https://aiwiki.ai), the free encyclopedia of artificial intelligence. Reuse freely with attribution.

**WeVisOne-VL** (WeChat Vision All-in-One) is a [vision-language model](https://aiwiki.ai/wiki/vision_language_model) family developed by the WeChat Vision team at [Tencent](https://aiwiki.ai/wiki/tencent). It combines object detection, language-guided localization, counting, pointing, text recognition, segmentation and image matting in one model. The released 2B and 4B variants build on [Qwen3-VL](https://aiwiki.ai/wiki/qwen3_vl), with a small additional decoder that produces masks and continuous opacity maps.[1][2][3]

The technical report was submitted to arXiv on October 22, 2026, and revised on October 24. The public release includes model weights, an inference package, demonstrations and evaluation code. Its focus is fine-grained image perception rather than general conversation.[1][3][4][5]

## Tasks and outputs

WeVisOne-VL expresses visual tasks through instructions and structured responses. Category names can select a class of objects, while referring expressions select objects by their attributes or relationships to other objects. The inference package exposes the following task families.[3]

| Task | Input query | Output |
| --- | --- | --- |
| Object detection | One or more category names | Bounding boxes for matching objects |
| Referring localization | A description of the target | Bounding boxes, or `None` when no target is predicted |
| Object counting | A category or description | Boxes or points, with a count |
| Object pointing | A category or description | Target points |
| Optical character recognition | A request to read or locate text | Recognized text and its bounding boxes |
| Referring segmentation | A category or description | Boxes and binary masks |
| Promptable matting | A category or description | Boxes and continuous alpha mattes |

Segmentation assigns pixels to an object or the background. Matting estimates fractional opacity, which can represent fine hair, soft boundaries and transparent regions. WeVisOne-VL predicts these dense outputs for individual targets identified by its language-and-coordinate response.[2][4]

## Architecture

The language model generates dedicated coordinate tokens from `<0>` to `<1000>` for normalized boxes and points. For segmentation and matting, it also generates an `<OBJ>` token after a predicted box. That token's hidden state supplies an instance representation to the pixel decoder.[2]

The decoder combines that representation with the predicted box and features from several levels of Qwen3-VL's vision encoder. It contains a visual feature adapter, a box-guided query decoder and an instance-conditioned dense prediction head. Matting adds refinement for detailed boundaries and fractional opacity. The complete decoder has about 9.6 million parameters; the released model cards specify 9,609,764 parameters in `mask_decoder.pt`.[2][4][5]

This decoder lets the deployed model produce masks and mattes without invoking a separate large segmentation model. That differs from the data-construction pipeline, which uses external perception models to create and check training annotations.[2]

## Training and annotation

The authors describe a data engine with an annotation inventory of more than 100 million images and approximately 3.6 billion bounding boxes. A knowledge graph organizes categories, aliases, relations and referring expressions. Complementary localization models propose candidate regions; a vision-language model verifies whether a region satisfies the query. Masks and alpha annotations retain the identity of the target associated with each box.[2]

The pipeline includes reviewed negative queries for which an image contains no matching object. A negative example is not established merely by changing a word in a positive query or obtaining an empty detector response: the authors describe checking the complete image before assigning an empty target set. Model errors subsequently guide annotation review and additional data collection.[2]

Training proceeds through a curriculum:[2]

| Phase | Main purpose |
| --- | --- |
| Initial supervised adaptation | Learn the added tokens, task instructions and response formats |
| Resolution and density scaling | Improve small-object recall and handle responses containing many objects |
| Token and pixel training | Connect localized instances to segmentation and matting predictions |
| Mixed-task reinforcement learning | Improve complete structured predictions using task-level rewards |

The reinforcement-learning phase uses group relative policy optimization. For segmentation and matting examples, its reward evaluates the predicted boxes rather than the masks or alpha mattes. Pixel quality also depends on the separately supervised decoder training.[2]

## Reported performance

The paper evaluates five object-level and two pixel-level task families across more than 30 benchmarks. Its reported inference protocol uses greedy decoding with temperature zero and a maximum of 1,960 visual tokens per image. The figures below come from the authors' evaluation, rather than an independent deployment study.[2]

| Benchmark and measure | 2B | 4B | Interpretation |
| --- | ---: | ---: | --- |
| LVIS, mean detection F1 | 53.3 | 55.3 | Higher is better |
| Dense200, mean detection F1 | 61.4 | 71.3 | Higher is better |
| RefCOCOg test, mean localization F1 | 78.2 | 81.8 | Higher is better |
| WeRef-Bench, absent-target rejection rate | 92.9% | 96.4% | Fraction of absent-target queries rejected |
| PixMo-Count test, exact-match accuracy | 88.1% | 89.6% | The predicted count must be exactly correct |
| TotalText, mean OCR F1 | 42.5 | 52.0 | Higher is better |
| ReasonSeg validation, generalized intersection over union | 60.9 | 71.2 | Higher is better |
| P3M-500-NP box matting, mean absolute difference multiplied by 1,000 | 6.6 | 7.3 | Lower is better |

For detection, referring localization and OCR, the reported mean F1 averages results across intersection-over-union thresholds from 0.5 to 0.95. These scores are not conventional detection average precision. Counting uses a different measure: an error of one object still counts as an incorrect answer.[2][3]

The 4B model's absent-target rejection is not perfect and varies by dataset. Its rate is 96.4% on WeRef-Bench, 67.4% on RefCOCOg-Rej and 52.6% on PR-Bench. A high score on one collection therefore does not establish that the model will reject every nonexistent target.[2]

The larger variant also does not win on every individual measure. The 2B variant has lower error on the P3M-500-NP box-matting test shown above. Task, dataset, prompt and evaluation measure all matter when comparing the models.[2]

## Using the released models

The official package loads the language-and-vision model together with its pixel decoder. Its model-card example constructs `WeVisOneVL` from either `tencent/WeVisOne-VL-2B` or `tencent/WeVisOne-VL-4B`, supplies an RGB image and calls `predict` with a task and query. The result includes a structured response and, for dense tasks, masks or alpha outputs.[3][4][5]

The repository documents Transformers and [vLLM](https://aiwiki.ai/wiki/vllm) backends. Its evaluation scripts use vLLM for text-and-coordinate tasks, and Transformers with the mask decoder for referring segmentation and matting. Evaluation also requires separately obtained benchmark images and annotations; the maintainers do not redistribute all evaluation data because of copyright restrictions.[3][6]

The package manifest requires Python 3.10 or later and pins its inference dependencies. The model cards identify `mask_decoder.pt` as a required component for the released dense-prediction architecture. Loading only a generic text-generation interface does not describe the complete documented pixel-output workflow.[4][5][7]

## Limitations and licensing

The authors identify a trade-off between perception specialization and general capabilities relative to the base model. The system directly predicts structured outputs without explicit chain-of-thought reasoning and currently targets single-turn image perception. The model cards state that the pixel decoder rejects video grids.[2][4][5]

Fine alpha boundaries remain sensitive to resolution, occlusion, transparency and prompt wording. The paper also distinguishes well-formed benchmark prompts from vague real-world instructions, which can make target selection more difficult.[2][4]

Both released model repositories declare Apache 2.0 in their license metadata. The code repository explicitly applies Apache 2.0 to Tencent-authored code and retains the original licenses and notices for third-party components.[3][4][5][8]

## References

1. Liang, Tianming, et al. [WeVisOne-VL: Unified Visual Perception from Objects to Pixels](https://arxiv.org/abs/2610.31329). arXiv:2610.31329, submitted October 22, 2026; version 2, October 24, 2026.
2. Liang, Tianming, et al. [WeVisOne-VL technical report, version 2](https://arxiv.org/html/2610.31329v2). Architecture, training, benchmark protocols, results and limitations.
3. Tencent WeChat Vision. [WeVisOne-VL repository and README](https://github.com/Tencent/WeVisOne-VL). Released resources, task interface and inference examples. Accessed November 4, 2026.
4. Tencent. [WeVisOne-VL-4B model card](https://huggingface.co/tencent/WeVisOne-VL-4B). Architecture, pixel decoder, intended use and limitations. Accessed November 4, 2026.
5. Tencent. [WeVisOne-VL-2B model card](https://huggingface.co/tencent/WeVisOne-VL-2B). Architecture and release information. Accessed November 4, 2026.
6. Tencent WeChat Vision. [Evaluation documentation](https://github.com/Tencent/WeVisOne-VL/blob/main/evaluation/README.md). Benchmark preparation, inference paths and measures. Accessed November 4, 2026.
7. Tencent WeChat Vision. [Package manifest](https://github.com/Tencent/WeVisOne-VL/blob/main/pyproject.toml). Python requirement and dependency versions. Accessed November 4, 2026.
8. Tencent WeChat Vision. [Apache License 2.0 and repository notices](https://github.com/Tencent/WeVisOne-VL/blob/main/LICENSE). Accessed November 4, 2026.