OneJev1.6-4B

OneJev1.6-4B

Reads a screenshot, photo, video or text with a few typed questions and returns a probability for each answer option in one pass. Built for developers of screen-driving agents.

Made by OmniJev 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: OneJev1.6-4B is a multimodal decision model built on Qwen3.5-4B. You give it a screenshot, photo, video or text plus a few typed questions, and it scores every answer option in one pass. It was fully fine-tuned on about 440,000 questions from agent runs, videos, images and public decision datasets, so it fits jobs like judging whether a task is done.

Who it suits: Developers building agents that read a screen and pick the next action are the main fit. Teams that need a quick yes or no on a finished task can use it too, and so can researchers who compare small decision models.

Industry | Who uses it | How they use it | Example
Software | Agent developers | Ask which action a screen agent should take next | Choose between click, type, scroll or stop from a screenshot
Operations | Automation teams | Check whether a task has finished | Ask whether an open invoice has been paid
Research | Model researchers | Compare small decision models on a shared board | Score it against other models of similar size on Decision Index

If you want long written answers or open chat, you can skip it.

Where to use it: Use it on any computer that runs Python.

• Windows, macOS or Linux with Python
• A local server started with the flide tool
• The Hugging Face transformers library

Use case | Who | What they get back
Screen agent steering | Agent developers | A probability for click, type, scroll and stop
Task completion check | Automation builders | A yes probability that the task is done
Progress tracking | Workflow teams | An expected level of how far along a task is

How to set it up: You need a computer with Python and some free disk space.

• Windows, macOS or Linux, with Python (the example uses Python 3.12)
• The uv tool, which the example uses to make its virtual environment
• Room for the weights, about 9.1 GB in BF16 or about 2.3 GB at 4-bit as an estimate

1. Install the Hugging Face hub package and download the model files if you want a local copy.
2. Create a Python 3.12 environment and activate it.
3. Install flide and torchvision in that environment.

pip install -U huggingface_hub
hf download OmniJev/OneJev1.6-4B

uv venv -p 3.12 && source .venv/bin/activate
uv pip install flide torchvision

How to use it: Start the server with the model name, then send a JSON request from another terminal. The request holds a state with the task and the screen, the media such as a screenshot, and your typed questions.

flide serve --model OmniJev/OneJev1.6-4B

Tool | What it does
noul question | Returns the probability of yes for a statement
choice question | Returns one probability for each option you list
score question | Returns the expected level on a scale you define

The request goes to the server's systemone endpoint. Answers come back under an answers key, one entry per question.

Example: You send a screenshot of an invoice page, the task of paying ACME's open invoice, and three questions. Is the task done, what should the agent do next among click, type, scroll or stop, and how far along is it. The server returns a yes probability for done, one probability for each next action, and an expected progress level.

For newcomers and experts: This model looks at a picture or a short video and answers your questions with probabilities instead of sentences. A good first try is to download the files with the two Hugging Face commands, start the server, and ask one yes or no question about a screenshot of your own.

Under the hood it is a full fine-tune of Qwen3.5-4B, about 4.5B parameters, that gives a calibrated probability per option in a single forward pass rather than writing text. One pass covers every typed question you attach. In return, you phrase each question as a yes or no check, a choice or a score.

Send a screenshot, photo, video or text
Add typed questions with their options
The model runs one forward pass
Read a probability for each option

Good to know: These scores are from Decision Index 0.3, using its public portion, and this model has the top overall score among those compared.

Title: Decision Index public part score
OneJev1.6-4B: 62.01 points
Perplexity Decider Flash v1.2: 60.06 points
CoCo-Decision-4B 2.3.0: 55.38 points
vLLM-SR d3-mini: 54.90 points
OneJev1.5-4B: 54.89 points
RSI-Jev v6.1 VL 4B: 50.98 points

• License: Apache 2.0
• Cost to run: none named. The weights take about 9.1 GB in BF16 or about 2.3 GB at 4-bit, as an estimate
• Data it reaches: the screenshots, photos, videos and text you send in a request. The source names no account or web service
• Not for: the source names no exclusions

On Hugging Face, 13 November 2026

• Task: image text to text
• Library: transformers
• Parameters: 4.5B
• Weights: about 9.1 GB in BF16; about 2.3 GB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 1,909
• Likes: 8
• License: apache-2.0
• Based on: Qwen/Qwen3.5-4B
• Quantized versions: 2
• Fine-tunes: 1
• Tags: transformers, safetensors, qwen3_5, image-text-to-text, onejev, system-one, decision-model, calibration, multimodal, gui-agent, video, conversational

Links

• Model: https://huggingface.co/OmniJev/OneJev1.6-4B
• Files: https://huggingface.co/OmniJev/OneJev1.6-4B/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download OmniJev/OneJev1.6-4B

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