Overview
Tencent WeChat Vision open-sourced WeVisOne-VL, with the technical report on arXiv from October 22, 2026 and weights, inference package, cookbooks, a Gradio demo and the evaluation pipeline released together under Apache 2.0 in two sizes, 2B (2.44B parameters) and 4B (4.83B parameters). The design folds object-level and pixel-level perception into one model: structured next-token prediction plus a lightweight 9.6M parameter pixel decoder produce masks and alpha mattes directly, so no external SAM is needed. Tasks cover object detection, referring localization, counting, pointing, OCR, referring segmentation and alpha matting, with granularity down to points, boxes, binary masks or continuous alpha. When a target is absent the model returns None instead of guessing a box.
Key Features
- Seven perception tasks in one model: Detection, referring localization, counting, pointing, OCR, referring segmentation and alpha matting share the same weights and the same inference path, so composite pipelines avoid shuttling data between specialized models.
- A 9.6M parameter pixel decoder: The hidden state of each object token, its predicted box and multi-level features from the Qwen3-VL vision tower are combined to decode full-resolution masks or alpha mattes, replacing the SAM-style segmentation model usually deployed alongside.
- Returns None for absent targets: Rejection rate on WeRef-Bench reaches 96.4 percent against 73.0 percent for the strongest open baseline, which removes a layer of manual fallback in review and labeling pipelines.
- Clear gains on dense detection: LVIS F1 mean reaches 55.3 (baseline 50.7), Dense200 hits 71.3 (baseline 58.7) and VisDrone reaches 44.7 (baseline 39.9).
- Usable OCR and counting: TotalText F1 mean is 52.0 (baseline 43.3) and PixMo-Count exact-match accuracy is 89.6 percent (baseline 82.8 percent), with CLOC at 28.3 percent and KubriCount at 40.1 percent.
- Matting is the standout: P3M-500-NP MAD sits at 7.3 (baseline 15.4) and referring matting on WeRefMat at 9.8 (baseline 22.4), one cascade stage fewer than stacking SAM2 with a matting model.
Use Cases
- Commerce and content review: detection, counting and OCR on product images in one pass, then filter by business rules.
- Image editing workflows: obtain an alpha matte directly for layer separation and background replacement, without a separate matting service.
- Documents and receipts: OCR localization plus structured output for scans with unstable layouts.
- Visual retrieval and referring: locate a target from one sentence of description and get an explicit empty result when it is missing.
- Data labeling: batch produce boxes, masks and mattes, then use the same model to screen out weak samples.
Pros
- Seven tasks in one model means pipelines no longer need SAM plus a separate detector, saving deployment work and VRAM.
- Rejection behavior is backed by checkable numbers, 96.4 percent on WeRef-Bench, more than twenty points above the open baseline.
- The 4B model beats 8B to 9B general vision-language models on dense detection, OCR and matting, a good size-to-quality ratio.
- Apache 2.0 licensing, with no extra restriction on commercial use or derivative work.
- Complete tooling: weights, inference package, cookbooks, Gradio demo and the evaluation pipeline are all published.
- Both 2B and 4B run on a single GPU, suitable for batch offline processing or edge deployment.
Pricing
Weights are public on Hugging Face and GitHub under Apache 2.0, allowing commercial use and further development. Local inference cost depends on your hardware; both the 2B and 4B variants run on a single GPU, and a hosted demo on Hugging Face is available to try.
Summary
WeChat Vision packed detection, OCR, segmentation and matting into one token-prediction model, swapped an external SAM for a 9.6M parameter decoder, and the 4B version pulls clearly ahead of 8B-class general vision models on dense detection and matting. The most practical part is rejection: returning None when a target is absent removes a layer of manual fallback in review and labeling. Teams building perception pipelines without wanting to maintain several specialized models should start here.
Version History
- Tencent WeChat Vision open-sources WeVisOne-VL 2B and 4B (2026-10-22): The technical report went up on arXiv on October 22 (v1) and was revised to v2 on October 24. Weights were released on Hugging Face and GitHub under Apache 2.0, 2.44B parameters for the 2B model and 4.83B for the 4B model, with a Qwen3-VL vision tower and a 9.6M parameter pixel decoder that outputs masks and alpha mattes. The inference package, cookbooks, a Gradio demo and the evaluation pipeline shipped at the same time; fine-tuning code is not out yet.