What is LocateAnything-3B?
LocateAnything-3B was published by NVIDIA on 26 May 2026 as part of its Eagle vision-language family. You give it an image and a phrase, such as "every forklift" or "the Save button", and it returns bounding boxes or points. It covers open-vocabulary object detection, referring expressions, crowded scenes, GUI elements, document layout and text location.
It pairs Moonshot's MoonViT-SO-400M vision encoder with Qwen2.5-3B-Instruct. Its main idea, Parallel Box Decoding, predicts each box's four coordinates in one step instead of one token at a time. It was trained on 12 million images with about 138 million queries and 785 million boxes. NVIDIA says the work also went into Nemotron 3 Nano Omni and Cosmos.
The weights use the NVIDIA License for academic and non-profit research only; commercial use is not permitted.
Is LocateAnything-3B worth it?
LocateAnything-3B is the fastest of the open grounding models NVIDIA compares it with, and its speed comes from a real design change rather than smaller weights. On one H100, NVIDIA measures 12.7 boxes per second against 5.0 for Rex-Omni and 1.1 for Qwen3-VL, the 2.5x and 10x gaps it advertises. The advantage grows with crowding, from 2x to 6x as the number of objects rises from 20 to 300.
Accuracy holds up too, on NVIDIA's own tests. It scores 60.3 mean F1 on ScreenSpot-Pro, a benchmark for finding buttons on professional software screens, and beats Rex-Omni on LVIS at the strictest overlap setting, 31.1 against 20.7. On document layout it leads by 6.1 points on DocLayNet and 14.5 on M6Doc. No independent leaderboard has repeated these results yet.
The licence decides most cases. NVIDIA allows academic and non-profit research only, and the language model inside carries Qwen's research licence as well. A robotics start-up or a labelling vendor can benchmark it but cannot put it in a product; Meta's SAM 3, which also finds objects from text, allows commercial use under its own licence.
Running it is easy for a 3B model. NVIDIA's 4K street-scene test on an A100 took about 8 seconds for a batch of four with 11.71GB of peak memory using its FlashAttention backend, against 35.12GB with standard attention. It runs through Transformers with remote code on Linux; TensorRT and Triton are not supported yet. It downloaded 204,296 times in the last month.
Where it earns its place
A university robotics lab
Find objects named in an instruction, such as "the red mug left of the sink", fast enough for a robot loop
Run Hybrid mode on one workstation GPU. At about 12 boxes per second on an H100 it keeps up with crowded scenes, and research use is what the licence allows.
A non-profit building an accessibility agent for desktop software
Locate the right button or field on a screen from a spoken request
Its 60.3 mean F1 on ScreenSpot-Pro is the reason to try it; pair it with a language model that decides what to click. Turning it into a paid product would need a different model.
A computer-vision PhD student
Auto-label a new detection dataset before human review
Batch inference with the la_flash backend handles 4K images in about 12GB. The labels serve the research; a commercial labelling service could not use them under the licence.
Before you commit
- Setup
- pip install the pinned packages (transformers 4.57.1 and others), load with trust_remote_code, and use generation_mode "hybrid" with max_new_tokens 8192; MagiAttention speeds it up on Hopper and Blackwell GPUs.
- Integrations
- Hugging Face Transformers with remote code, a batch inference script, NVIDIA's Hugging Face Space demo. TensorRT, TensorRT-LLM and Triton not yet supported.
- Data and privacy
- Self-hosted, images stay local. NVIDIA License: academic and non-profit research only, no commercial use; built on Qwen2.5-3B-Instruct under the Qwen Research License and MoonViT under MIT.
- Cost at scale
- Free to run for research; fits on one 24GB or larger GPU. Commercial deployment is not licensed at any price NVIDIA publishes.
- Limits
- Images up to about 2.5K in production use, prompts up to 24K tokens, up to 8,192 new tokens per answer; Linux only; benchmark figures self-reported.
What’s good
- 12.7 boxes per second on one H100, 2.5 times Rex-Omni
- 60.3 mean F1 on ScreenSpot-Pro for GUI grounding
- Boxes and points for objects, text, layout and screen elements in one model
- About 12GB peak memory on a 4K image with the FlashAttention path
What’s not
- Academic and non-profit research only; no commercial use
- Benchmarks are NVIDIA's own, not independently repeated
- Linux and Transformers only; no TensorRT or Triton support
- Outputs boxes and points, not masks
Best for
University and non-profit labs that need fast open-vocabulary detection or GUI grounding for research.
LocateAnything-3B alternatives
LocateAnything-3B: common questions
- Is LocateAnything-3B worth it?
- LocateAnything-3B finds things in pictures faster than the open models NVIDIA tested it against, and NVIDIA will not let you ship it. We rate it 3.8 out of 5.
- How much does LocateAnything-3B cost?
- Weights free for academic and non-profit research; commercial use not permitted. No hosted API.
- What are the downsides of LocateAnything-3B?
- Academic and non-profit research only; no commercial use; benchmarks are NVIDIA's own, not independently repeated; linux and Transformers only; no TensorRT or Triton support; outputs boxes and points, not masks.
- Who is LocateAnything-3B best for?
- University and non-profit labs that need fast open-vocabulary detection or GUI grounding for research.
- How much does LocateAnything-3B cost for a team or in production?
- Free to run for research; fits on one 24GB or larger GPU. Commercial deployment is not licensed at any price NVIDIA publishes.
- Is LocateAnything-3B safe for company data?
- Self-hosted, images stay local. NVIDIA License: academic and non-profit research only, no commercial use; built on Qwen2.5-3B-Instruct under the Qwen Research License and MoonViT under MIT.
- What are the best LocateAnything-3B alternatives?
- The closest alternatives we have reviewed are SAM 3 (Segment Anything Model 3), Nemotron 3.5 Super VL and MiniCPM-V 4.7 (35B-A3B and 1B).
Answers checked Nov 7, 2026.