Breaking

NVIDIA Releases Open-Source Detection VLM LocateAnything-3B, 10x Faster

June 29, 2026 at 08:15 EDT

Foundation Models
Open Source
Research & Papers

NVIDIA Research has open-sourced LocateAnything, a unified vision-language model (VLM) framework for object detection and visual grounding. Its new Parallel Box Decoding (PBD) method parallelizes the bounding-box coordinate generation that previously bottlenecked VLMs, lifting detection throughput by up to roughly 10x over prior approaches.

May 26, 2026 · NVIDIA Research

LocateAnything-3B: detect anything you can name — at ~10× the speed

An open-sourced 3B vision-language model that predicts every bounding box in a single parallel step. The result: dense-scene detection without the autoregressive bottleneck — and without losing accuracy.

12.7

Boxes/sec on a single H100 (Hybrid mode)

+3.8%

LVIS mean F1 vs Rex-Omni — faster and more accurate

Parameters — runs on consumer GPUs

Throughput on a single H100 (Boxes Per Second)

Each block ≈ 1 BPS. Taller column = more boxes per second.

1.1

Qwen3-VL

5.0

Rex-Omni

12.7

LocateAnything-3B

Over 10× Qwen3-VL and about 2.5× Rex-Omni.

Why it's faster: Parallel Box Decoding

Conventional VLM — one token at a time

x₁→ y₁→ x₂→ y₂→ …

PBD — every box as one atomic unit, in parallel

box 1 box 2 box 3 ↘ all at once ↙

In dense scenes scaling from 20 → 300 objects, a 2×–6× speedup was observed.

One model, six localization tasks

60.3

ScreenSpot-Pro (GUI) — SOTA

76.8

DocLayNet (layout)

70.1

M6Doc (layout)

43.3

TotalText (OCR)

+1.8%

COCO mean F1

2.5K

Native input resolution

Trained on 12M images · 138M queries · 785M bounding boxes — MoonViT encoder paired with Qwen2.5-3B-Instruct.

What developers like

Lightweight 3B yet fast and accurate; runs on consumer GPUs
Practical for GUI navigation and dense detection
Python-free C++/ggml community port at high speed
Useful for dataset labeling

Continue reading

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