BREAKING
NVIDIA Open-Sources LocateAnything-3B
Parallel Box Decoding Replaces Token-by-Token
1Old: sequential tokens
2PBD: one parallel step
3Lower latency
12.7 BPS on a Single H100
LocateAnything-3B12.7
Rex-Omni5
Qwen3-VL1.1
0%
LVIS F1 vs Rex-Omni
0%
COCO mean F1
0
ScreenSpot-Pro SOTA
One Model, Many Tasks
Capabilities
Open-vocab detection
GUI and OCR grounding
Document layout parsing
Notes
Research-only license
Eagle VLM family
Runs on consumer GPUs
Lightweight, Fast Detection Goes Open
AI NEWS BLITZ
NVIDIA has released a 3-billion-parameter detection model that locates anything from text.