NVIDIA releases LocateAnything-3B vision-language model with 2.5× faster object detection via parallel box decoding
NVIDIA released LocateAnything-3B, a 3-billion parameter vision-language model that predicts bounding boxes in parallel rather than token-by-token, achieving up to 2.5× higher throughput compared to autoregressive approaches. The model, trained on 12M images with 138M+ queries and 785M bounding boxes, supports object detection, GUI element grounding, and robotics perception.
LocateAnything-3B — Quick Specs
NVIDIA releases LocateAnything-3B vision-language model with 2.5× faster object detection via parallel box decoding
NVIDIA released LocateAnything-3B, a 3-billion parameter vision-language model designed for fast visual grounding tasks including object detection, GUI element localization, and robotics perception.
Technical specifications
The model uses Qwen2.5-3B-Instruct as its language backbone and MoonViT-SO-400M as its vision encoder. Training data spans 12 million images, 138 million queries, and 785 million bounding boxes across natural scenes, robotics, driving, GUI interaction, and document understanding domains.
Input supports images up to 2.5K resolution and text prompts up to 24K tokens. The model can generate up to 8,192 new tokens during inference and uses BF16 precision with KV cache.
Parallel Box Decoding innovation
LocateAnything's core technical contribution is Parallel Box Decoding (PBD), which predicts complete bounding box coordinates in a single parallel step instead of generating coordinates token-by-token autoregressively. According to NVIDIA, this approach delivers up to 2.5× higher throughput while preserving geometric consistency.
The architecture outputs structured coordinate tokens in fixed-length blocks of 6 tokens, including semantic labels, box coordinates, and control tokens. The model supports three decoding modes: Fast Mode (parallel prediction), Slow Mode (autoregressive), and Hybrid Mode (parallel with autoregressive fallback).
Supported use cases
NVIDIA lists these supported applications:
- Open-set and long-tail object detection
- Dense multi-object detection in cluttered scenes
- Phrase and referring-expression grounding
- Automated dataset labeling
- GUI element grounding for agentic systems
- Robotics and autonomous driving perception
- Document understanding and OCR localization
- Industrial inspection and surveillance
The model has been integrated into NVIDIA's Nemotron 3 Nano Omni production models for grounding and multimodal capabilities.
Availability and licensing
LocateAnything-3B is available on Hugging Face and GitHub as of May 26, 2026. The model is released under NVIDIA's non-commercial license, permitting use only for academic and non-profit research. Commercial use is prohibited except by NVIDIA and its affiliates.
The model runs on NVIDIA Ampere, Hopper, Blackwell, and Lovelace architectures. TensorRT and TensorRT-LLM support is not yet available. Deployment on embedded platforms like NVIDIA Thor requires additional optimization including quantization or distillation.
What this means
Parallel Box Decoding represents a practical efficiency gain for visual grounding tasks that require detecting multiple objects or UI elements in a single image. The 2.5× throughput improvement matters for real-time applications like robotics and GUI automation where latency compounds across multiple API calls. However, the non-commercial license limits this to research settings, and the model's integration into production Nemotron systems suggests NVIDIA views visual grounding as infrastructure for agentic AI systems rather than standalone capability. The 3B parameter size makes local deployment feasible on consumer GPUs while the multi-domain training data indicates this is positioned as a generalist localization model rather than domain-specific detector.
Related Articles
Nvidia Claims Groq 3 LPX Hits 3,400 Tokens/Sec, 4x Cerebras — But Needs 64 Chips to Do It
Nvidia's new Groq 3 LPX inference accelerator hit 3,400 tokens per second on Gemma 4 31B, a figure the company says is four times faster than Cerebras. Experts note the benchmark uses at least 64 LPX chips versus Cerebras' one or two accelerators, making the comparison far less clean than it appears.
SenseNova Releases U1.5-8B-MoT, an Open-Weight Unified Model for Image Generation and Editing
SenseNova has released SenseNova-U1.5-8B-MoT, an open-weight native multimodal model built on its NEO-unify architecture for image generation, editing, and native 4K output. The model is available on Hugging Face under an Apache 2.0 license, with no inference pricing yet since it must be self-hosted.
Nvidia Research: Custom Harness Pushes Claude Opus 5 from 30% to 100% on ARC-AGI-3
Nvidia researchers report that a custom agent harness with a supervisory component drove Claude Opus 5 from a 30% baseline score to a perfect 100% on the ARC-AGI-3 benchmark. The findings add to mounting evidence that scaffolding, not the underlying model, is the dominant factor in long-horizon agentic performance.
DeepSeek Releases Experimental V4-Flash-Vision-Exp, Claims Near-Parity With Opus 4.8 on Agent Benchmarks
DeepSeek has released V4-Flash-Vision-Exp, an experimental multimodal extension of V4-Flash that adds image understanding while preserving text reasoning capabilities. The company claims the model approaches or beats Anthropic's Opus 4.8 on its internal multimodal agent benchmarks.
Comments
Loading...