Tencent Releases Open WeVisOne-VL Model Combining Detection, OCR and Segmentation

The model comes in two sizes, listed on Hugging Face at 2.44 billion and 4.83 billion parameters, with weights under the Apache 2.0 licence. It is built on Alibaba's Qwen3-VL. The pitch to developers is that one model can do the work that usually needs a separate object detector, OCR engine and segmentation tool.

How It Works

WeVisOne-VL treats every task as structured next-token prediction. It adds coordinate tokens for normalised boxes and points. When the object asked about is not in the image, the model is designed to answer "None" rather than guess a box.

For pixel-level output, the model emits a special token after each predicted box. A 9.6-million-parameter decoder then turns that token's internal state into a full-resolution mask or matte. The decoder reuses features from the Qwen3-VL vision encoder, so it does not need a second image encoder or Meta's Segment Anything model.

Tencent built the training data with a knowledge-graph-driven engine that annotated more than 100 million images with about 3.6 billion boxes, MarkTechPost reported. Training ran in three supervised stages, followed by a reinforcement learning stage on 178,000 filtered examples.

Benchmark Claims

All the performance figures are self-reported by Tencent. The company says the 4B model beats similar-sized vision-language models on most of more than 30 benchmarks. On the Dense200 dense-detection test it scored 71.3 F1, against 58.7 for Nvidia's LocateAnything-3B. It reached 81.8 F1 on the RefCOCOg referring test and 52.0 F1 on the TotalText OCR benchmark, compared with 25.7 for the specialist PP-OCRv5.

The weakest result shows where the model still struggles. On PR-Bench it rejected only 52.6 per cent of queries about objects that were not there, so in many such cases it still drew a box. On Tencent's own rejection benchmark it reported 96.4 per cent.

Availability

The model runs on a single GPU through the Transformers or vLLM libraries, and the package also lists NPU and CPU devices. Tencent has not published VRAM requirements or the context length for deployment, and fine-tuning code is not yet released. Community-made GGUF versions can do object-level tasks but cannot produce masks, because the decoder needs the model's native visual features. Tencent has also open-sourced Mask2Alpha, the matting annotation tool from its data pipeline.

Why It Matters

For companies that run document processing, retail shelf analysis, inspection or image-editing systems, small open models that combine several perception tasks could reduce how many models they need to deploy and maintain. The Apache 2.0 licence also allows commercial use, which LocateAnything-3B's non-commercial licence does not. Independent testing will show whether Tencent's own benchmark gains hold up.