Tencent's open WeVisOne-VL finds, counts and cuts out objects in images

Tencent’s WeChat Vision team has open-sourced WeVisOne-VL, a vision-language model in 2B and 4B parameter sizes that handles object detection, OCR, counting, pointing, segmentation and alpha matting in a single model. The weights are published under the Apache 2.0 licence, and the 4B model’s Hugging Face page was created on October 22, 2026.

The pitch is consolidation. Cutting an object out of a photo usually means chaining a detector, a segmentation model and a matting model, each with its own coordinate convention and confidence threshold; WeVisOne-VL produces boxes, points, counts, text locations, masks and alpha mattes from one natural-language prompt. Every benchmark result attached to the release comes from the authors’ own evaluation, and none had been independently reproduced as of late October.

What Tencent released with WeVisOne-VL

The name stands for “WeChat Vision All-in-One”. The project README describes it as a vision-language foundation model that unifies two families of tasks:

• Object-level perception: object detection, referring localization (finding the object a phrase describes, such as “the man in the red jacket”), counting, pointing and OCR.
• Pixel-level perception: referring segmentation and promptable alpha matting.

An alpha matte records how much each pixel belongs to an object, as a value between 0 and 1, rather than the hard yes-or-no of a binary mask. That matters wherever a boundary is not a clean edge, such as hair, fur, motion blur, smoke and glass.

The October release bundle, per the repository, includes the technical report, the 2B and 4B model weights, an inference package, cookbooks and an evaluation pipeline. Tencent also open-sourced Mask2Alpha, the matting annotator its data engine used to turn binary masks into alpha mattes, on a separate branch, and a demo is available for both sizes. Fine-tuning code appears on the project’s to-do list.

How WeVisOne-VL goes beyond boxes and points

Earlier grounding models (models that tie words to places in an image) such as Kosmos-2, Florence-2, Rex-Omni and LocateAnything stop at boxes and points, according to an engineering review by Miraflow; getting a mask usually meant bolting on a separate segmentation model, and a clean cutout meant adding a matting model on top. WeVisOne-VL keeps the Qwen3-VL backbone and its next-token interface, then adds binary masks and continuous alpha mattes.

It does this through “structured next-token prediction” plus a lightweight pixel decoder. The interface works like this, per Miraflow’s reading of the release:

• 1,001 coordinate tokens, from <0> to <1000>, with four tokens per box and two per point.
• A “None” output when the target is absent.
• An <OBJ> token after each instance that needs pixel output, which hands off to the decoder.

The pixel decoder has about 9.6 million parameters and outputs masks and alpha mattes with no SAM (Meta’s Segment Anything Model) required. Miraflow puts the exact count at 9,609,764 parameters, about 0.2% of the 4B model, and says the decoder reuses the VLM’s own vision-transformer features and refines alpha in a cascade at quarter, half and full resolution.

The released Python package exposes these capabilities through a task argument with values including detection, referring, point, count_box, count_point, ocr, ocr_grounding, segmentation and matting.

Training data

Tencent built the training set with a knowledge-graph-driven data engine that discovers concepts, localizes them with complementary expert models verified by a VLM, and extends the verified boxes to masks and alpha mattes. Miraflow reports the scale as about 100 million annotated images, with supervised fine-tuning stages of 90.2M, 40.5M and 41.8M samples, followed by GRPO (a reinforcement-learning method) on 178K uncertainty-selected samples.

WeVisOne-VL specs at a glance

Spec	WeVisOne-VL-2B	WeVisOne-VL-4B
Parameters	2,441,744,384	4,830,581,248
Licence	Apache 2.0
Backbone	Qwen3-VL
Pixel decoder	About 9.6M parameters, no SAM
Checkpoint files (4B)	Not stated	9.66 GB model.safetensors plus 38.5 MB mask_decoder.pt
Library	Transformers, Safetensors
Visual tokens in evaluation	Up to 1,960 per image, greedy decoding

Tencent’s benchmark claims, and what they leave open

The README compares WeVisOne-VL-4B against the strongest open baseline on each benchmark, drawn from Qwen3-VL and Qwen3.5 models up to 9B, Rex-Omni, LocateAnything and task-specific segmentation and matting models. In Tencent’s own evaluation, the headline numbers are:

Task	Benchmark (metric)	Best baseline	WeVisOne-VL-4B
Detection	LVIS (F1 mean, higher is better)	50.7	55.3
Dense detection	Dense200 (F1 mean)	58.7	71.3
Referring	RefCOCOg test (F1 mean)	77.6	81.8
Rejection	WeRef-Bench (rejection rate)	73.0	96.4
Counting	PixMo-Count test (accuracy)	82.8	89.6
Pointing	PointBench (accuracy)	69.9	71.9
OCR	TotalText (F1 mean)	43.3	52.0
Referring segmentation	ReasonSeg val (gIoU)	65.0	71.2
Box matting	P3M-500-NP (MAD, lower is better)	15.4	7.3
Text-prompted matting	WeRefMat referring (MAD)	22.4	9.8

The full tables show a more mixed picture on some tests. On Dense200 pointing, LocateAnything-3B scores 87.6 against 87.0 for the 4B model. On HumanRef referring detection, Rex-Omni-3B’s 79.9 beats the 2B model’s 77.4, though the 4B reaches 80.7. On RefCOCOg referring segmentation, UniPixel-7B scores 77.1, above the 2B’s 74.9 and just below the 4B’s 77.8. In box matting on P3M-500-NP, the smaller 2B model scores better than the 4B, at 6.6 against 7.3.

The OCR comparison includes a dedicated specialist: PP-OCRv5 scores 58.6 on SROIE, while WeVisOne-VL-2B scores 62.1 and the 4B scores 68.1 in Tencent’s tables.

Two caveats apply. First, the rejection benchmark WeRef-Bench and the matting benchmark WeRefMat carry Tencent’s own “We” naming, and the reported results on them are Tencent’s. Second, as Miraflow notes, every number comes from the authors’ paper and README and none has been independently reproduced yet. Tencent points to its evaluation pipeline for reproducing the public benchmarks.

Where to get WeVisOne-VL and how to run it

The 4B model is on Hugging Face under the Apache 2.0 licence, ungated, tagged as an image-text-to-text model built on the qwen3_vl architecture and loadable with the Transformers library. The page links an arXiv technical report numbered 2610.31329. At the time of the snapshot, the 4B page showed 11,255 downloads and 98 likes, and listed no hosted inference providers.

The code, inference package, cookbooks and evaluation scripts live in Tencent’s GitHub repository, which ships an English and a Simplified Chinese README. The 4B repository ships the main weights and the mask decoder as separate files that the inference package loads as one engine, so there is no second checkpoint to version.

Apache 2.0 is a permissive licence that allows commercial use, modification and redistribution, which makes the release usable in products without the custom terms that attach to some open-weight models.

What WeVisOne-VL means for builders

The practical case is fewer hand-offs. A pipeline that today chains a detector, a segmenter and a matting model can, in principle, send one text query that selects the instance and produces its matte. For product listings, thumbnails or inpainting masks, that removes the coordinate conversions and per-stage thresholds where cutouts quietly fail.

The rejection behaviour may matter as much as the cutouts. Tencent says the model returns None for absent targets instead of hallucinating. That lets a script refuse to make a cutout rather than return a plausible wrong one. In Tencent’s referring-detection tables, competitors’ rejection rates on RefCOCOg range from 0.5% for Rex-Omni-3B to 22.5% for Qwen3-VL-8B, against 67.4% for WeVisOne-VL-4B.

Limits to plan around

• Not small next to SAM. The 4B checkpoint has about 4.8 billion parameters, against 224.4 million for SAM 2.1 Large; the value is integration and language grounding, not a lighter footprint.
• Edge quality varies. Alpha edges are sensitive to resolution, occlusion, transparency and prompt phrasing.
• Single-turn, no reasoning trace. The model is single-turn with no chain of thought.
• No fine-tuning code yet. Fine-tuning code is still listed among the project’s to-do items.
• Self-reported results. Until outside teams rerun the benchmarks, the comparisons rest on Tencent’s numbers.

What to watch next for WeVisOne-VL

The first thing to watch is independent reproduction of the benchmark tables, particularly the rejection and matting results that lean on Tencent’s own WeRef-Bench and WeRefMat sets. The second is the fine-tuning code on the repository’s to-do list, which would let teams adapt the model to their own object categories. The third is whether hosted inference providers pick up the model; the Hugging Face page listed none at the time of the snapshot. Finally, the open-sourced Mask2Alpha annotator gives other teams a way to build their own alpha-matte training data, which may show up in future matting models.