Interactive installations have needed the same thing for twenty years: tell me what’s in front of the camera. For most of that time the answer was a fixed list — face, body, hand, or a model trained on 80 preset categories that will never include the object your piece is actually about.
Open-vocabulary detection removes the list. You describe what you want in plain language — “the red umbrella,” “a person holding a cup,” “anything made of glass” — and the model returns boxes. mudler/locate-anything.cpp-gguf does that with no Python at inference time, and our tracker caught it up 55% in a day to 1.42 million downloads.
What it is
A C++/ggml inference engine — the same lineage as llama.cpp — running NVIDIA’s LocateAnything-3B (itself built on Qwen2.5-3B). Two capabilities:
- Open-vocabulary detection: find objects matching arbitrary text, not a fixed class list
- Visual grounding: given a natural-language query, localise the region it refers to
Builds and sizes:
| Build | Size | Note |
|---|---|---|
| F16 | 9.15GB | full precision |
| Q8_0 | 6.26GB | reported box-identical to f32, recommended |
| Q6_K | 5.51GB | also reported box-identical |
| Q4_K | 4.72GB | smallest, minor precision trade-offs |
The maintainer reports 3.9–5.5× speedups over the official PyTorch implementation, tested on a Ryzen 9 9950X3D CPU and NVIDIA GB10, with identical detections. Usage is a single CLI call with a model, an image and a prompt. Weights carry NVIDIA’s own licence; the GGUF conversion is MIT. Check the former before shipping anything commercial.
One honest note on timing: this repository was created in June. What’s new is the adoption curve, not the release — which is the same pattern we’ve been tracking all week, where packaging rather than publication is what moves the numbers.
Why it matters for installation work
Three practical reasons this beats a cloud vision API for gallery pieces:
No Python at runtime. Anyone who has maintained an installation knows the failure mode: it worked in June, and in October a dependency updated and the environment is broken. A single compiled binary plus a model file is the difference between a piece that runs for a year and one that needs a visit.
No network. Camera frames never leave the machine. That removes the consent problem, the bandwidth problem, and the “the venue’s wifi went down” problem in one move.
It runs on a CPU. Not fast enough for every frame, but fine for the pattern most installations actually need — check the scene once or twice a second and react.
The obvious build is a piece that responds to whatever a visitor brings into the room, described in words you wrote rather than categories someone else chose. That has been possible with cloud APIs for a few years, and locally, on ordinary hardware, for about five minutes.