NVIDIA's LocateAnything-3B, released in July 2026, is a 3-billion-parameter open-vocabulary vision-language model (VLM) designed to precisely locate specific objects, text, or user interface (UI) elements within an image using bounding box coordinates. Built on a highly integrated architecture combining the MoonViT-SO-400M visual encoder and the Qwen2.5-3B-Instruct language model, it instantly infers the visual location of a target specified by the user in natural language. By merging the language model's exceptional contextual understanding with high-resolution image processing capabilities, it goes beyond simple object detection, excelling in tasks such as providing visual grounding that aligns with complex descriptive sentences and inferring the location of GUI elements. Existing large-scale vision-language models have inherent structural limitations, including slow inference speeds and high computational resource consumption, due to their adoption of an autoregressive approach that sequentially predicts text tokens to derive the coordinates of a target. To overcome this bottleneck, LocateAnything-3B introduces a parallel box decoding (PBD) method. Similar to how a high-performance ultrasound diagnostic device scans the entire contrast profile of biological tissue at once, it extracts a large number of location area coordinates at ultra-high speed with just a single parallel computation step. Despite its lightweight 3B architecture, the parallel computation mechanism enables significantly faster processing compared to conventional autoregressive models. It also fully supports ComfyUI custom nodes and ggml C++ porting, providing excellent operational accessibility, allowing it to run quickly and easily in general research environments or on local workstations. In biomedical and biotechnology research, LocateAnything-3B serves as a powerful automation tool for detecting fine structures such as cells, protein aggregates, and fluorescent staining areas in microscopic images or histology slide images using only natural language queries. When a researcher inputs a somewhat complex referring expression, such as "the green fluorescent apoptotic area around the cell nucleus" or "mitochondrial aggregates within the cytoplasm," the model instantly specifies the precise coordinates of the target object without requiring fine-tuning. This enables efficient pre-screening by automatically selecting regions of interest (ROI) in high-resolution bio-image datasets in real-time and connecting them to subsequent quantitative analysis pipelines. Furthermore, it plays a crucial role in building automated bio-labs and automated control systems for laboratory equipment. It can accurately infer the location of reagent container labels, pipette tip boxes, and control buttons on equipment control software in real-time camera footage or software interface screenshots, and then pass the precise coordinates to a robotic arm or control script. In the process of virtualizing the experimental environment and automating robotic control, it functions as a visual detection and location tracking (UI/Object Grounding) engine, maximizing experimental reproducibility and research productivity.

<!

Automated Segmentation of Microstructures and Extraction of ROI Regions in Cell Microscopy Images: Given a high-resolution fluorescence/brightfield cell microscopy image (1024x1024 or higher), the system receives a specific descriptive phrase such as "apoptotic cells marked with green fluorescence" or "mitochondrial aggregates within the cytoplasm." Within 1-2 seconds, it generates precise bounding box coordinates and passes them to a quantitative analysis pipeline.