NVIDIA · 2026-05-26 · major
NVIDIA LocateAnything-3B — Parallel Box Decoding Vision-Language Grounder Hits 12.7 Boxes/Sec on H100, ~10× Faster Than Qwen3-VL, Trained on 138M Queries Across 785M Boxes
A 3B vision-language grounding model that emits whole bounding boxes in parallel instead of token-by-token. Apache 2.0 code, NVIDIA-licensed weights, 138M-query training set, and a live HF Space demo.
LocateAnything decodes full bounding boxes in one shot, getting NVIDIA a vision-language grounder that is roughly 10× faster than Qwen3-VL.
Key specs
| Parameters | 3B |
|---|---|
| Throughput | 12.7 boxes/s on H100 |
| Vs qwen3 vl | ~10× faster |
| Training queries | 138M |
| Training boxes | 785M |
| Lvis f1 | 50.7 |
| Screenspot pro mean f1 | 60.3 |
What is it?
A 3B vision-language model from NVIDIA Labs for visual grounding: open-set detection, dense detection, referring-expression grounding, GUI element grounding, OCR localization, and document layout. Built on a MoonViT vision encoder plus Qwen2.5-3B-Instruct language model.
How does it work?
The novelty is Parallel Box Decoding (PBD), which treats a bounding box as an atomic unit and predicts its four coordinates in one forward pass instead of serializing coordinates into a token stream. Three inference modes (fast / slow / hybrid) trade off speed against robustness. Trained on LocateAnything-Data: 12M images, 138M queries, 785M boxes, labelled by humans plus Qwen3-VL, Molmo, and SAM 3.
Why does it matter?
Autoregressive VLMs cap throughput because every coordinate is another decoding step. PBD turns localization into a parallel-friendly task, unlocking real-time grounding for robotics, GUI agents, and document AI without giving up accuracy. Hits LVIS F1@Mean 50.7 and ScreenSpot-Pro 60.3 mean F1 — competitive with much larger models.
Who is it for?
Researchers and product teams building visual grounding into GUI agents, robotics perception, autonomous driving, or document understanding.
Try it
pip install transformers, then load `nvidia/LocateAnything-3B` and call `generate(..., generation_mode="hybrid")`.