SandLogicTechnologies/WeVisOne-VL-4B-GGUF

WeVisOne-VL-4B

WeVisOne-VL-4B is a unified vision-language model developed by Tencent's WeChat Vision team for visual perception tasks ranging from object-level understanding to pixel-level image processing. Built upon Qwen3-VL, the model combines visual reasoning with structured perception capabilities within a single multimodal architecture.

The model brings together open-vocabulary object detection, referring localization, object counting, object pointing, and OCR with pixel-level capabilities such as referring segmentation and promptable image matting. Its design uses structured next-token prediction together with a lightweight pixel decoder to support both semantic and dense visual perception.

The GGUF variants in this repository provide reduced-size representations of WeVisOne-VL-4B for local inference and experimentation while preserving the underlying multimodal model architecture.

---

Model Overview

• Model Name: WeVisOne-VL-4B
• Base Model: Qwen3-VL
• Parameter Count: Approximately 5 Billion Parameters
• Architecture: Unified Vision-Language Model
• Modalities: Image and Text
• Vision Capabilities: Object Detection, Localization, Counting, Pointing, OCR, Segmentation, Matting
• Developer: Tencent / WeChat Vision
• License: Apache 2.0

WeVisOne-VL is designed to unify object-level and pixel-level visual perception. It uses dedicated coordinate tokens to represent bounding boxes and points, while a lightweight pixel decoder handles dense visual outputs such as referring segmentation and full-resolution alpha mattes.

---

Quantization Formats

This repository provides multiple GGUF quantized versions of WeVisOne-VL-4B designed to reduce storage and memory requirements. The available formats offer different trade-offs between model footprint and numerical precision.

IQ4_NL

• Size reduction of approx. 70.35% (2.44 GB) compared to 16-bit (8.23 GB)
• 4-bit non-linear quantization providing a compact representation of the multimodal model weights
• Reduces the memory footprint substantially while retaining more weight information than the supplied 3-bit variant
• Suitable for visual perception workloads involving OCR, object detection, localization, counting, and image understanding
• Provides a practical balance between model compression and numerical representation for local multimodal deployment

IQ4_XS

• Size reduction of approx. 71.81% (2.32 GB) compared to 16-bit (8.23 GB)
• Compact 4-bit quantization intended for deployments where storage and memory requirements need to remain low
• Provides a smaller representation than IQ4_NL while remaining within the same 4-bit quantization class
• Suitable for local experimentation with visual grounding, OCR, object-level perception, and other image-based tasks
• Offers an alternative when reducing the GGUF footprint is more important than retaining the slightly larger IQ4_NL representation

IQ3_M

• Size reduction of approx. 75.82% (1.99 GB) compared to 16-bit (8.23 GB)
• Aggressive 3-bit quantization providing the smallest GGUF representation among the supplied variants
• Significantly lowers storage and memory requirements for constrained local environments
• Suitable for lightweight multimodal perception experiments where minimizing model size is a primary requirement
• The lower numerical precision can introduce a greater trade-off for detailed visual perception and pixel-sensitive tasks

Q6_K

• Size reduction of approx. 58.93% (3.38 GB) compared to 16-bit (8.23 GB)
• Higher-precision 6-bit K-Quant representation among the supplied quantized variants
• Retains more numerical information than the lower-bit conversions while remaining substantially smaller than the F16 model
• Suitable for visual workloads where maintaining a higher-precision weight representation is preferred
• Provides the largest quantized representation in this repository and can be considered for more demanding visual perception workloads

---

Training Background (Original Model)

WeVisOne-VL-4B is built upon Qwen3-VL and extends the vision-language architecture toward a unified visual perception system. Instead of separating object-level and pixel-level tasks into independent models, the architecture is designed to handle multiple perception tasks through structured output prediction and a dedicated pixel-processing component.

Unified Visual Perception

• Supports open-vocabulary object detection for identifying visual entities
• Performs referring-expression localization and object pointing using structured coordinate outputs
• Supports object counting for visual scenes
• Provides OCR capabilities for extracting textual information from images
• Combines object-level perception with pixel-level visual processing

Pixel-Level Processing

• Uses a lightweight pixel decoder containing approximately 9.6 million parameters
• Processes multi-level visual features produced by the Qwen3-VL vision component
• Uses a box-guided query decoder for dense visual prediction
• Supports referring segmentation based on visual targets and textual references
• Produces full-resolution continuous alpha mattes for image matting tasks

Structured Visual Outputs

• Uses dedicated coordinate tokens ranging from <0> to <1000> for normalized spatial representations
• Can represent bounding boxes and points through structured token sequences
• Produces a special instance token for dense object-related predictions
• Returns structured outputs that can be consumed by downstream visual-processing systems
• Uses a unified model to cover multiple object-level and pixel-level perception tasks

---

Key Capabilities

• Object Detection
  Identifies visual objects using open-vocabulary detection capabilities rather than being restricted to a fixed set of predefined categories.

• Visual Grounding
  Connects textual descriptions or referring expressions with corresponding regions or locations within an image.

• Object Counting
  Determines the number of relevant visual instances in an image based on the supplied task or query.

• OCR and Text Localization
  Detects and interprets text within visual inputs while supporting localization of relevant textual regions.

• Object Pointing
  Represents target locations through structured point coordinates for visual perception workflows.

• Referring Segmentation
  Produces pixel-level segmentation masks corresponding to objects or regions described by a textual reference.

• Image Matting
  Supports promptable alpha matting and generates continuous alpha information for separating foreground subjects from their background.

• Unified Visual Perception
  Combines object-level and pixel-level perception capabilities within a single vision-language model.

---

Usage Example

./llama-cli \
  -m SandlogicTechnologies/WeVisOne-VL-4B-IQ4_NL.gguf \
  --mmproj WeVisOne-VL-4B.mmproj-f16.gguf \
  -p "Describe the important visual information in this input."

Recommended Usecases

• Visual Grounding Applications
  Build systems that associate natural-language descriptions with specific objects or regions within images.

• Object Detection and Counting
  Develop visual analytics workflows requiring open-vocabulary detection and instance counting.

• OCR and Text Localization
  Process images containing text and identify relevant textual regions for downstream document or visual analysis.

• Visual Search and Reference Resolution
  Locate objects based on textual descriptions and use structured coordinates for downstream processing.

• Image Segmentation
  Create applications that isolate objects or regions according to referring expressions.

• Image Matting
  Build foreground extraction and background-removal workflows using promptable alpha-matting capabilities.

• Unified Visual Perception Research
  Experiment with combining detection, localization, counting, OCR, segmentation, and matting within a common vision-language architecture.

• Local Multimodal Inference
  Use the GGUF variants when the original F16 representation requires more storage or memory than the target deployment environment provides.

WeVisOne-VL-4B is intended primarily for research and development involving visual perception and multimodal understanding. The official model documentation notes that it is not intended for autonomous high-impact decisions or individual identification.

---

Acknowledgments

These quantized models are based on the original work by the Tencent / WeChat Vision development team.

Special thanks to:

• The Tencent WeChat Vision team for developing and releasing WeVisOne-VL and its unified visual perception architecture.

• The open-source multimodal research community for contributing to the tooling and ecosystem surrounding vision-language models and visual perception.

• Georgi Gerganov and the llama.cpp open-source community for enabling efficient GGUF model conversion and local inference tooling.

---

Contact

For questions, feedback, or support, please reach out at support@sandlogic.com or visit https://www.sandlogic.com/.