WeVisOne-VL-4B

A person can use it to detect and locate objects, count or point to them, perform OCR localization, segment image regions, and create alpha mattes.

Availability

Model checkpoint available at tencent/WeVisOne-VL-4B; inference package installable from GitHub

Pricing model

Not stated

Deployment

Inference example loads the model with device="auto"

Privacy

Not established

Licence

Tencent-authored code is Apache License 2.0; third-party components retain their original licenses

• The pixel decoder outputs masks and alpha mattes without requiring SAM. ^[1]

  - 🪶 **Lightweight Pixel Decoder**: a 9.6M-parameter decoder outputs masks and alpha mattes, with no SAM required.

• Model weights, an inference package, cookbooks, and an evaluation pipeline are released. ^[1]

  - **[2026-10]** 🎉 We release the technical report, model weights (2B & 4B), inference package, Cookbooks, and evaluation pipeline!

• A 4B demo is available on Hugging Face Spaces. ^[1]

  - [x] Model weights (2B & 4B) on Hugging Face

• The source reports WeVisOne-VL-4B achieves a 96.4% rejection rate on WeRef-Bench. ^[1]
• The model supports open-vocabulary object detection, referring localization, object counting, object pointing, and OCR. ^{[1] [2]}

  Built upon Qwen3-VL, it unifies object-level perception (open-vocabulary object detection, referring localization, object counting, object pointing, and OCR) and pixel-level perception (referring segmentation and promptable alpha matting) through structured next-token prediction and a lightweight 9.6M-parameter pixel decoder.

• The model supports referring segmentation and promptable alpha matting. ^{[1] [2]}

  Built upon Qwen3-VL, it unifies object-level perception (open-vocabulary object detection, referring localization, object counting, object pointing, and OCR) and pixel-level perception (referring segmentation and promptable alpha matting) through structured next-token prediction and a lightweight 9.6M-parameter pixel decoder.

• The release includes the tencent/WeVisOne-VL-4B model and a required mask_decoder.pt component. ^[2]
• The model targets single-turn image perception, and video grids are rejected by the pixel decoder. ^[2]

  - For inference efficiency, the model directly generates structured outputs without explicit chain-of-thought (CoT) reasoning, and currently targets single-turn image perception (video grids are rejected by the pixel decoder).

• The release is intended for research and is not intended for autonomous high-impact decisions or identifying individuals. ^[2]

  The release is intended for research in visual grounding, referring localization, pointing, counting, OCR localization, referring segmentation, and image matting.

• A speech model adds multilingual transcription and translation Daily brief 2026-10-26