LocateAnything-3B

NVIDIA's 3B vision-language model that returns boxes and points for objects, GUI elements and text, up to 2.5x faster, research-only.

In short. LocateAnything looks at an image and a short request, such as 'the red cup' or 'all cars', and answers with the position of each match as a box or a point. It writes each box in one step instead of one number at a time, which makes it faster, but NVIDIA releases it for research only.

Summary

LocateAnything-3B is NVIDIA's vision-language model for visual grounding. Given an image and a text prompt, it returns labels with bounding boxes or points, for open-set detection, dense detection, referring expressions, GUI elements, document layout and text localization. It combines a MoonViT vision encoder, a Qwen2.5-3B-Instruct language model and an MLP projector, for 3.83B parameters in the checkpoint. Its core idea, Parallel Box Decoding, predicts each complete box or point as one unit in a single step instead of writing coordinates token by token, and it offers Fast, Hybrid (the default) and Slow modes. Training used 12M images and more than 138M queries from natural scenes, robotics, driving, GUI and documents, and the card says NVIDIA built the model into Nemotron 3 Nano Omni.

NVIDIA's project page reports 12.7 boxes per second on one H100 in the default Hybrid mode, against 1.1 for Qwen3-VL and 5.0 for Rex-Omni. It reports mean F1 gains of 3.8 points on LVIS and 1.8 on COCO over Rex-Omni at the same model size, 60.3 mean F1 on ScreenSpot-Pro for GUI grounding, 76.8 on DocLayNet, 70.1 on M6Doc, 43.3 on TotalText and 78.7 on HumanRef. All of these are NVIDIA's own results, and the paper is arXiv 2605.27365.

NVIDIA releases the model for research and development only. The NVIDIA License allows use for research or evaluation and lets only NVIDIA and its affiliates use it commercially, and the Qwen2.5-3B-Instruct language model inside carries the Qwen Research License. The card shows inference with Transformers in BF16 on Linux and says TensorRT, TensorRT-LLM and Triton are not yet supported. It accepts images up to 2.5K, prompts up to 24K tokens and up to 8,192 new tokens. The release date on the card is 26 May 2026. In the 30 days to 11 November 2026, Hugging Face counted about 112,000 downloads, and the repository has more than 3,100 likes.

Good for

  • Research on grounding, detection and GUI agents that need fast, precise boxes and points from a text prompt
  • Pre-labeling images with boxes or points for research datasets, from category names or referring phrases
  • Comparing grounding speed and accuracy against Rex-Omni and Qwen3-VL

Watch out for

  • The license is research or evaluation only, commercial use is reserved for NVIDIA, and the Qwen2.5 language model inside has its own research license
  • Every speed and accuracy figure comes from NVIDIA, and the speed is measured on an H100
  • It runs only through Transformers on an NVIDIA GPU under Linux, and TensorRT, TensorRT-LLM and Triton are not yet supported
  • The output is structured coordinate tokens that you must parse, and Hybrid mode falls back to slower decoding when a block looks irregular

Performance claims come from the developer and have not been independently verified.

How much memory does it need?

Estimates for the LocateAnything-3B variant. A mixture-of-experts model must hold all of its experts in memory even though only some are used for each token, so total parameters decide the memory and active parameters decide the speed.

Variants Parameters BF16 FP8 / INT8 4-bit (Q4) Fits on
LocateAnything-3B BF16 3.83B 8.81 GB 4.4 GB 2.64 GB Laptop or phone

Estimates: weights plus about 15% overhead. Real usage grows with context length and depends on the software.

VRAM calculator

The license in plain words

NVIDIA's short research-only license, as shipped with LocateAnything-3B. You may use, copy, modify and share the work, but only non-commercially, which the text defines as research or evaluation purposes only. NVIDIA and its affiliates keep the right to use it commercially.

More about NVIDIA Non-Commercial

See also

Frequently asked questions

What is LocateAnything-3B?
LocateAnything looks at an image and a short request, such as 'the red cup' or 'all cars', and answers with the position of each match as a box or a point. It writes each box in one step instead of one number at a time, which makes it faster, but NVIDIA releases it for research only.
Can I use LocateAnything-3B commercially?
Not without separate permission. LocateAnything-3B is released under the NVIDIA Non-Commercial, which allows research and personal use only.
Is LocateAnything-3B open source?
LocateAnything-3B is open-weight, not open source in the strict sense. You can download the model files, but the NVIDIA Non-Commercial adds conditions that the Open Source Initiative definition does not allow.
Can I run LocateAnything-3B on my own computer?
Yes. The smallest variant, LocateAnything-3B, needs about 2.64 GB of memory at 4-bit precision, which fits a modern laptop or a mid-range graphics card.
When was LocateAnything-3B released?
LocateAnything-3B was released by NVIDIA on May 26, 2026.

Official links

Sources

Specifications come from these primary sources. Benchmark claims, where mentioned, are reported by the developer.

Release date and facts last verified November 11, 2026.

Related models

Model Released Parameters Context License Smallest variant needs
Nemotron 3 Nano Omni NVIDIA 31B 3B active 256K NVIDIA Open Model (Permissive) 32 GB Gaming GPU
Qwen3.6 Alibaba Qwen 27.8B – 36B 3B active 256K Apache 2.0 (Permissive) 19 GB Gaming GPU
Gemma 4 Google DeepMind 5.1B – 30.7B 256K Apache 2.0 (Permissive) 3.52 GB Laptop or phone
Mage-VL Microsoft 4.74B 256K Apache 2.0 (Permissive) 3.27 GB Laptop or phone