LocateAnything-3B-GGUF

LocateAnything-3B GGUF (for llama.cpp)

GGUF model files for running nvidia/LocateAnything-3B in llama.cpp's MTMD (multimodal) support.

LocateAnything-3B is a grounding / open-vocabulary detection VLM: a MoonViT-SO-400M vision encoder

• an Eagle-MLP connector + a Qwen2.5-3B text decoder. Given an image and a description it returns bounding boxes as <ref>label</ref><box><x1><y1><x2><y2></box>, where each coordinate is a control token <0>..<1000> (pixels = coord/1000 * image_dim).

llama.cpp support:

• PR1 - converter + model support, next-token (slow) decode: #24749

A parallel box-decoding "fast mode" (--grounding-mode fast|hybrid) is in development on top of PR1.

These models are only compatible with the PR branch, not upstream llama.cpp master.

---

Download

huggingface-cli download sabafallah/LocateAnything-3B-GGUF --include "locateanything-3b-bf16.gguf" --local-dir gguf_models/locateanything
huggingface-cli download sabafallah/LocateAnything-3B-GGUF --include "mmproj-locateanything-3b-bf16.gguf" --local-dir gguf_models/locateanything

---

Run example

llama-mtmd-cli (the included dog.jpg is a sample input):

build/bin/llama-mtmd-cli \
  -m gguf_models/locateanything/locateanything-3b-bf16.gguf \
  --mmproj gguf_models/locateanything/mmproj-locateanything-3b-bf16.gguf \
  --image dog.jpg --chat-template chatml -ngl 99 --temp 0 -n 64 \
  -sys "You are a helpful assistant." \
  -p "<image 1><__media__>Locate all the instances that matches the following description: dog."

Output: <ref>dog</ref><box><131><54><795><986></box>.

Files