Vela 3.0 · Studio

Vela 3.0 · Studio

Answers questions about an image or a text with a small 0.25B encoder running on CPU: point at an object, extract fields, catch OCR mistakes or decide between options. For developers trying the model.

Made by vllm-sr on Hugging Face. These notes are Mindket's, written from its Space page and README; the app's own files and license are on Hugging Face.

About

What it is: Vela 3.0 Studio is a page for trying the vela3-fast-vision model. You give it a picture or some text, and the model answers in real time on CPU. The model is an encoder, so it answers questions about what you give it and does not write text. Each question runs on one of three backbones built on UltraVBERT.

Who it suits: Developers and researchers who want to try a small vision and text model for document checks, and people who need quick answers about an image.

Industry | Who uses it | How they use it | Example
Document processing | Developers | Pull fields out of scanned invoices into structured data | Running an invoice image through the Extract tab
Finance operations | Reviewers | Check OCR output against the image it came from | Highlighting characters the image does not support
Research | Vision researchers | Probe how a small encoder handles pointing and decisions | Naming an object in a picture and reading the box

You can skip it if you want a model that writes or chats, since this one only answers questions.

Where to use it:

• A web browser, on the Space's own page at Hugging Face
• A script or tool that calls the app's API, which the page shows in curl and Python form
• A copy on your own computer, using Docker on Windows, macOS or Linux

Use case | Who | What they get back
Point at an object | Anyone with a picture | One box around the named thing and its strongest cell
Extract from images | Developers | The fields and values of a document as JSON
Catch OCR mistakes | Reviewers | The characters the image does not support, highlighted

How to set it up:

• A browser, if you only use the Space online
• Docker and Python with pip, if you run a copy yourself

1. Open the Space's page on Hugging Face to use it online. Nothing needs installing.
2. For a local copy, install the huggingface_hub package and download the Space's files.
3. Move into the new folder and build the Docker image.
4. Run the image with port 7860 open.

pip install -U huggingface_hub
hf download vllm-sr/vela3-studio --repo-type space --local-dir vela3-studio
cd vela3-studio
docker build -t vela3-studio .
docker run -it -p 7860:7860 vela3-studio

How to use it: Pick a tab, add an image or a text, and ask your question. Each tab ships with saved example outputs that load at once, and a separate option runs your own input through the model.

Tool | What it does
Point at it | Finds an object you name in a picture and returns one box
Extract from images | Turns an image into fields and values as JSON
Catch OCR mistakes | Highlights characters in OCR output that the image does not support
Decide | Gives a probability for each option of a choice question about an image
Text spans | Finds personal data, named entities and claims a source text does not support

Example: You upload a scan and ask "What type of document is this?" with a few options listed. The Decide tab gives you a probability for each option.

For newcomers and experts: If you are new to this, open the Point at it tab and name something in one of the pictures. You get a box and a heat map showing where the model looked.

For experts, a distilled decision backbone answers choice and yes or no questions. A second one carries the box adapter. The untouched original backbone holds the adapters for OCR checks, image hallucination and text. Each question costs one 0.25B pass. Tesseract reads the text for extraction, and the source says any OCR engine would do.

Upload an image or paste a text
Pick a tab and ask a question
One of three backbones answers it
The answer shows on the page

Good to know:

• License: apache-2
• Cost to run: none named, and it runs on Hugging Face's basic CPU hardware
• Data it reaches: the images and text you give it, and the text Tesseract reads from images
• Not for: writing or generating text, since it only answers questions
• The source reports these scores from its own runs, not as board entries

Title: Region pointing, mean IoU on BoxSet-test v2
Vela 3.0 fast vision: 44.1 IoU
A 27B development build: 44.2 IoU

On Hugging Face, 13 November 2026

• Likes: 9
• Built with: docker
• License: apache-2.0

Links

• Space: https://huggingface.co/spaces/vllm-sr/vela3-studio
• Files: https://huggingface.co/spaces/vllm-sr/vela3-studio/tree/main

Try it

Open the Space in your browser, or copy it to your own Hugging Face account from its page. Nothing in this kit runs by itself.