VLX-Seek-1.5-10B

VLX-Seek-1.5-10B

Finds and points at objects in images by choosing among candidate regions, and answers None when the target is absent. For drone, robot and surveillance camera vision work.

Made by omlab on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: VLX-Seek-1.5-10B is a 10B image and text model from omlab for finding and pointing at objects in a picture. It is built for drones, robots, robotic dogs, surveillance cameras and inspection systems. Instead of writing box coordinates as numbers, it picks from candidate regions of the image, and it can answer that a target is not there.

Who it suits: It suits teams that need a model to tell what is in a scene and which exact instance you mean. The card describes research and development use in fine-grained perception.

Industry | Who uses it | How they use it | Example
Robotics | Robot and robotic dog developers | Add region level perception to a robot's vision module | Locating the right cup among several in the robot's view
Drones | Drone vision teams | Find small objects in dense scenes seen from above | Counting vehicles in a drone frame
Security | Surveillance system builders | Ground a described object in camera frames | Finding the person a description refers to
Inspection | Inspection system teams | Check whether a target is present or absent | Getting None back when the target is missing

You can skip it if you only need image captions, since it is aimed at answers tied to regions of the image.

Where to use it:

• Python with the transformers library, on Windows, macOS or Linux
• Drones, robots and robotic dogs
• Surveillance cameras and inspection systems
• Edge side devices or a robot base station

Use case | Who | What they get back
Open vocabulary detection | Vision researchers | Located regions for the objects you name
Referring expressions | Robotics developers | The one region that matches a description
Object counting | Drone analysts | Counts backed by region level evidence
Absent target checks | Inspection teams | An explicit None when nothing matches

How to set it up:

• Python and pip
• Disk space: roughly 20 GB for the BF16 files, or roughly 5 GB for a 4-bit copy (estimated)
• The project's GitHub repository from omlab, which holds the inference code

1. Install the Hugging Face hub package.
2. Download the model files to your computer.
3. Open the om-ai-lab VLX-Seek repository on GitHub and follow its steps for installation and model loading.

pip install -U huggingface_hub
hf download omlab/VLX-Seek-1.5-10B

How to use it: The model runs through a custom pipeline with four stages, and the card sends you to the repository for the code and runnable examples. That is where the install steps, inference scripts, prompt templates and visualization examples are.

Tool | What it does
Region proposals | Generates candidate regions in the image
Region tokens | Turns each region into something the language model can refer to
Prompt formatting | Builds the prompt that carries your question and the regions
Post-processing | Converts the model's output into the final answer

Example: Show it a camera frame and ask for the person in the red coat. It refers to the matching region, or answers None when nobody in the frame fits.

For newcomers and experts: You give it a picture and ask where something is. It answers by pointing at one of the regions it found, or it says None. For a first try, download the files with the two commands above and read the example scripts in the repository.

Most vision language models write bounding box numbers out as text. This one turns candidate regions into addressable tokens and answers by selecting, comparing and referring to them, which fits what language models do well. Version 1.5 adds a stronger auxiliary vision tower, faster OPN proposal generation, more Linear Attention layers for speed and memory, and hard negative training for the None output. Omlab plans three sizes for the 1.5 family, 0.6B, 3B and 10B, and this page covers the largest.

An image comes in with your question
Candidate regions are proposed
Regions become addressable tokens
The model selects and refers to regions
You get region answers or None

Good to know: The model works in Chinese and English. The card shows its results as pictured tables and a blog post, so no scores are quoted here.

• License: Apache 2.0
• Cost to run: no API key or paid service is named, you supply the hardware
• Data it reaches: downloads model files from Hugging Face and works on the images you give it
• Not for: work outside fine-grained perception, since the card lists only research and development uses in that area

On Hugging Face, 13 November 2026

• Task: image text to text
• Library: transformers
• Parameters: 10B
• Weights: about 20 GB in BF16; about 5 GB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 137,772
• Likes: 39
• License: apache-2.0
• Languages: zh, en
• Quantized versions: 1
• Tags: transformers, safetensors, vlx_seek_1_5, text-generation, object-detection, grounding, multimodal, zero-shot-object-detection, image-text-to-text, conversational

Links

• Model: https://huggingface.co/omlab/VLX-Seek-1.5-10B
• Files: https://huggingface.co/omlab/VLX-Seek-1.5-10B/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download omlab/VLX-Seek-1.5-10B

Then follow the usage in its model card. Nothing in this kit runs by itself.