LocateAnything-3B-GGUF

LocateAnything-3B

LocateAnything-3B is a specialized vision-language model developed by NVIDIA for visual grounding and object localization tasks. This repository contains GGUF quantized variants of the model optimized for efficient local inference.

Unlike general-purpose multimodal assistants, LocateAnything-3B is designed to identify the precise spatial location of objects, regions, text, GUI elements, and entities referenced in natural language prompts. The model combines strong visual understanding with high-quality localization capabilities, enabling efficient deployment across perception-oriented workflows.

The model introduces Parallel Box Decoding (PBD), which predicts complete geometric structures simultaneously rather than autoregressively generating coordinate tokens, resulting in improved throughput and localization quality.

---

Model Overview

• Model Name: LocateAnything-3B
• Base Model: nvidia/LocateAnything-3B
• Architecture: Vision-Language Model with Parallel Box Decoding
• Parameter Count: 3 Billion
• Modalities: Text, Image
• Primary Languages: English
• Developer: NVIDIA
• License: nvidia-license
• Core Capability: General-purpose visual grounding and localization

---

Quantization Formats

This repository provides various GGUF quantized versions of the LocateAnything-3B model optimized for efficient multimodal inference and visual grounding workloads.

IQ3_M

• Size reduction of approx 76.18% (1.51 GB) compared to 16-bit (6.34 GB)
• Aggressive 3-bit quantization optimized for maximum memory efficiency
• Suitable for low-memory multimodal deployment environments
• Enables practical execution of visual grounding models on consumer hardware
• Fine-grained localization precision and dense grounding performance may reduce compared to higher-precision variants

IQ4_NL

• Size reduction of approx 70.35% (1.88 GB) compared to 16-bit (6.34 GB)
• Advanced 4-bit non-linear quantization designed to better preserve localization fidelity and multimodal understanding quality
• Better suited for grounding workflows requiring stronger coordinate consistency and object localization accuracy
• Designed to reduce quantization loss compared to more aggressive formats
• May require slightly increased computational overhead during inference

IQ4_XS

• Size reduction of approx 71.77% (1.79 GB) compared to 16-bit (6.34 GB)
• Balanced 4-bit quantization focused on efficient inference and dependable visual grounding performance
• Provides a practical balance between memory efficiency, localization quality, and runtime speed
• Suitable for general-purpose grounding applications and multimodal deployment scenarios
• Maintains stable performance across most practical localization workloads

---

Training Background (Original Model)

LocateAnything-3B is trained with a focus on large-scale visual grounding, multimodal localization, and spatial reasoning across diverse visual domains.

Pretraining

• Trained using large-scale multimodal datasets spanning natural scenes, documents, robotics, driving environments, and GUI interfaces
• Focus on aligning visual representations with natural language descriptions
• Optimized for precise object localization and grounding tasks

Grounding Optimization

• Enhanced using large-scale grounding datasets containing millions of localization annotations
• Introduces Parallel Box Decoding to improve throughput and geometric consistency
• Optimized for referring expression grounding, dense localization, and multimodal perception workflows

---

Key Capabilities

• Visual Grounding Localizes objects and entities referenced through natural language descriptions.

• Referring Expression Grounding Identifies precise regions corresponding to textual descriptions within images.

• Multi-Object Localization Supports grounding multiple entities within complex visual scenes.

• GUI Element Grounding Locates interface components and screen elements in GUI environments.

• Text Localization Supports grounding of textual regions embedded within visual inputs.

• Efficient Local Deployment Quantized variants enable practical multimodal inference on consumer hardware.

---

Usage Example

Using llama.cpp

./llama-mtmd-cli \
  -m SandlogicTechnologies/LocateAnything-3B_IQ4_NL.gguf \
  --mmproj SandlogicTechnologies/mmproj-LocateAnything-3B-BF16.gguf \
  --image image.png \
  -p "Analyze the diagram and explain the underlying concept step-by-step."

Recommended Usecases

• Visual Grounding Systems Build applications capable of locating objects from natural language descriptions.

• Multimodal Perception Pipelines Integrate visual localization capabilities into perception-oriented AI systems.

• GUI Understanding Workflows Support interface automation and GUI element identification tasks.

• Document and Scene Understanding Ground textual descriptions within complex visual environments.

• Research and Experimentation Evaluate multimodal localization techniques and grounding workflows.

---

Acknowledgments

These quantized models are based on the original work by the NVIDIA Research team.

Special thanks to:

• The NVIDIA team for developing and releasing the LocateAnything-3B model.
• Georgi Gerganov and the llama.cpp open-source community for enabling efficient GGUF quantization and inference.

---

Contact

For questions, feedback, or support, please reach out at [email protected] or visit https://www.sandlogic.com/