Tencent’s WeChat Vision team has released WeVisOne-VL, an open vision language model in 2B and 4B sizes. The WeVisOne-VL vision language model handles detection, referring, counting, pointing, OCR, segmentation and alpha matting in one model. Pixel outputs come from a 9.6M-parameter decoder, so no separate SAM is needed. It also answers “None” when a queried object is absent. For developers, that can replace a detector, an OCR engine and a segmenter with one model.
TL;DR
- Size: 2B (2.44B params) and 4B (4.83B params) on Hugging Face, and a 9.6M-parameter pixel decoder. Trained with up to 8,192 tokens of context; deployment context length is not disclosed.
- Runs on: A single GPU through Transformers or vLLM in BF16. The package also lists NPU and CPU devices. VRAM needs are not disclosed. A community GGUF ran object-level tasks on a 6 GB RTX 2060.
- Performance: Beats the same-scale VLM baselines in Tencent’s tables on most of 30+ benchmarks (self-reported).
- Best: 71.3 F1 Mean on Dense200 dense detection, versus 58.7 for LocateAnything-3B.
- Worst: 52.6% rejection rate on PR-Bench, so many absent-target queries still get a box.
What is WeVisOne-VL?
WeVisOne-VL (“WeChat Vision All-in-One”) is a perception-focused model from Tencent built on Qwen3-VL. It targets fine-grained localization rather than open-ended chat. Users choose the output granularity: points, boxes, binary masks or continuous alpha mattes.
How does WeVisOne-VL work?
Every task is cast as structured next-token prediction. The model adds coordinate tokens <0> to <1000> for normalized boxes and points. An absent target produces None instead of a guessed box.
For pixel tasks, the model emits an
- Visual Feature Adapter: reuses multi-level Qwen3-VL ViT features instead of a second image encoder.
- Box-Guided Query Decoder: 6 deformable cross-attention blocks seeded by the
- Dense Prediction Head: refines output at S/4, S/2 and S/1 into a full-resolution mask or matte.
How was it trained?
A knowledge-graph-driven data engine annotated over 100M images with about 3.6B boxes. SAM3, Rex-Omni and an in-house detector proposed regions, and Qwen3-VL verified them.
Training used 3 SFT stages. Stage 1 aligned output formats on 90M examples. Stage 2 raised input to 1,960 visual tokens and up to 500 boxes per response. Stage 3 trained the pixel decoder, then specialized it for matting on 130M samples. A final GRPO stage used 178K uncertainty-filtered examples, as described in the research paper.
How does WeVisOne-VL perform on benchmarks?
All numbers below are self-reported, using greedy decoding and at most 1,960 visual tokens per image. Detection, referring and OCR use F1 averaged over IoU 0.5 to 0.95.
- Detection: 55.3 F1 on LVIS and 71.3 on Dense200 for the 4B model.
- Referring: 81.8 F1 on RefCOCOg test, versus 77.6 for LocateAnything-3B.
- Rejection: 96.4% on Tencent’s in-house WeRef-Bench (2,880 queries), versus 73.0%.
- Counting: 89.6% exact-match on PixMo-Count test, versus 82.8%.
- OCR: 52.0 F1 on TotalText, ahead of the PP-OCRv5 specialist at 25.7.
- Segmentation: 71.2 gIoU on ReasonSeg val, versus 65.0 for LISA-13B.
- Matting: 7.3 MAD on P3M-500-NP box prompts, versus 15.4 for SAM2-Matting (SAM3). Lower is better.
The 2B model trails the 4B model on most tasks. It still scores better box-prompted matting on P3M-500-NP, at 6.6 MAD.
‘});h+=’
‘+(x.lo?’Lower is better (MAD ×10³).’:’Higher is better.’)+’
‘;var el=R.querySelector(‘#wvo-bars’);el.innerHTML=h;setTimeout(function(){el.querySelectorAll(‘.fill’).forEach(function(f){f.style.width=Math.max(f.dataset.w,1.2)+’%’})},40);setTimeout(post,60)}
drawBars(0);
window.addEventListener(‘load’,post);window.addEventListener(‘resize’,post);setTimeout(post,300);
})();
