BREAKING
NVIDIA Open-Sources LocateAnything-3B
0
B
parameters
0
BPS
on one H100
Parallel Box Decoding
1
Old: x1 then y1 then x2 then y2
↓
2
New: predict full box at once
↓
3
Faster and more accurate
Speed vs Rivals
vs Qwen3-VL
10
vs Rex-Omni
2.5
0
%
LVIS F1 gain
0
high-IoU vs 20.7
Research-Only, Runs on RTX 4090
AI NEWS BLITZ
NVIDIA Research has open-sourced a 3-billion-parameter vision-language object detector.