Skip to content

LocateAnything 3B

nvidia/LocateAnything-3B
Open Source · chat · open-weights
Open Alert me on changes
Context
Max output
Weights
Open
API $/1M
Modalities
text · image
Released
02 Mar 2026
License: other · nvidia/LocateAnything-3B
Download image Share on X Share on LinkedIn
AI summary
● machine-written

LocateAnything-3B: NVIDIA vision-language model for visual grounding

LocateAnything-3B is a 3 billion parameter vision-language model from NVIDIA that performs visual grounding tasks including open vocabulary object detection, text localization, and target pointing in images and videos. It uses Parallel Box Decoding (PBD) to predict bounding boxes and keypoints as complete structures in parallel, improving both accuracy and inference speed. The model is built on Qwen2.5-3B for language and MoonViT for vision processing, trained on a dataset of over 138 million samples across natural scenes, robotics, autonomous driving, GUI interaction, and document understanding.

What's new
  • Implements Parallel Box Decoding (PBD) for block-wise multi-token coordinate prediction
  • Trained on LocateAnything-Data dataset with 138+ million samples across multiple domains
  • Supports multiple visual grounding tasks: object detection, OCR, GUI element localization, and video target pointing
  • Available in multiple quantized formats (BF16, Q4_K_M, Q5_K_M, Q6_K, Q8_0) via GGUF
Best for
Open vocabulary object detection in images and videosDocument understanding and OCR text localizationGUI agent automation and element localizationRobotics and autonomous driving perception tasks
Sources

Source: https://huggingface.co/nvidia/LocateAnything-3B