Breaking

NVIDIA Releases LocateAnything-3B Detection VLM With Up to 10x Faster Parallel Prediction

June 28, 2026 at 15:15 EDT

Foundation Models
Open Source
Research & Papers

NVIDIA Research has open-sourced LocateAnything-3B, a 3-billion-parameter Vision-Language Model that predicts an object's full bounding box in a single parallel step, making object detection up to 10x faster than prior VLMs while reportedly improving accuracy rather than reducing it.

NVIDIA Research · Open Source

LocateAnything-3B: detecting objects faster by deleting one step

A ~3B vision-language detector that predicts each bounding box in a single parallel step — substantially speeding up inference while accuracy goes up, not down.

12.7

boxes / sec on a single H100 (Hybrid Mode)

~3B

parameters · runs on RTX 4090

785M

bounding boxes across 12M training images

Inference speed vs other VLM detectors

Relative throughput · taller column = faster

10×+

LocateAnything
vs Qwen3-VL

2.5×

LocateAnything
vs Rex-Omni

1×

baseline
detector

Gains widen in dense scenes — scaling 2–6× from 20 to 300 boxes per image.

Parallel Box Decoding (PBD)

Conventional

Autoregressive, token by token

x1→ y1→ x2→ y2

Latency stacks up with every object.

PBD

One box as an atomic unit

x1 · y1 · x2 · y2 → one step

Predicted in a single parallel step.

Accuracy gains vs Rex-Omni

+3.8%

LVIS mean F1

+1.8%

COCO

31.1 vs 20.7

at strict high-IoU (0.95)

What testers liked

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.