OneJev1.6-0.8B

OneJev1.6-0.8B

Answers typed questions about a screenshot, photo, video or text with a probability for each option. It suits developers who need fast decisions in an agent loop.

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: OneJev1.6-0.8B makes quick decisions from visual or text input. It belongs to the OneJev 1.6 family of multimodal decision models and has 0.8B parameters. It averages the weights of three fine-tunes of Qwen3.5-0.8B. The input can be text or material such as screenshots, photos and video clips, with a few questions you type. It replies with a probability for every option in one pass.

Who it suits: It suits developers building agents that must decide what to do next, check whether a task is done, or judge progress from what is on screen.

Industry | Who uses it | How they use it | Example
Software | Agent developers | Ask what an agent should do next from a screenshot | Choosing between click, type, scroll and stop
Automation | Workflow builders | Check whether a task has finished | Asking if an invoice has been paid
Research | Model evaluators | Compare small decision models on a public index | Reading the model's Decision Index score

People who want free-form chat answers can skip it, since it returns probabilities for options you define.

Where to use it:

• A Python 3.12 environment on Linux or macOS, using the card's commands
• A local server started with the flide tool
• Agent loops that send screenshots and read back probabilities

Use case | Who | What they get back
Next action choice | Agent developers | One probability for each action option you list
Task completion check | Automation builders | A probability that the task is finished
Progress scoring | Agent developers | An expected progress level on a scale you set

How to set it up:

• Python 3.12 and the uv tool
• The flide and torchvision packages
• About 1.7 GB of disk for the BF16 weights

1. Create a Python 3.12 virtual environment and activate it.
2. Install flide and torchvision.
3. Start the local server with this model.

uv venv -p 3.12 && source .venv/bin/activate
uv pip install flide torchvision
flide serve --model OmniJev/OneJev1.6-0.8B

How to use it: In a second terminal, send a JSON request to the local server's systemone endpoint. The request holds a state with a task and a screen image, the media itself, and a set of typed questions. The card shows a curl example that encodes a screenshot as base64.

Tool | What it does
noul question | Asks a yes or no question and returns the probability of yes
choice question | Returns one probability per option you describe
score question | Returns the expected level on a scale you list

Example: You send a screenshot with the task of paying an open invoice. You ask whether the invoice is paid, what the agent should do next, and how far along the task is. The server returns the yes probability, one probability per action and an expected progress level.

For newcomers and experts: A decision model does not write sentences. It picks among options you give it and says how sure it is. A first small step is to start the server and send the card's example with a screenshot of your own.

Training used between 390,000 and 440,000 questions that came from agent runs, videos, images and public decision datasets. On the maker's benchmark board, Decision Index 0.3, it scores 48.82, the best result among models up to 2.5B parameters.

Screenshot, photo, video or text
Typed questions with their options
One forward pass through the model
Probability for every option

Good to know:

• License: Apache 2.0.
• Cost to run: none named. The steps need no API key or paid service.
• Data it reaches: it downloads files from the Hugging Face Hub. Screenshots you send stay on your local server.
• Not for: free-form text answers. The model returns probabilities for options you define.
• The Decision Index scores are the maker's report, run with the board's own kit.

Title: Decision Index public part score
OneJev1.6-0.8B: 48.82 points
OneJev1.5-0.8B: 42.71 points
vLLM-SR d3-nano: 42.22 points
Sieve 2B v2: 37.26 points
vLLM-SR d3-lite: 36.93 points
Crivo-150M: 36.39 points

On Hugging Face, 13 November 2026

• Task: image text to text
• Library: transformers
• Parameters: 853M
• Weights: about 1.7 GB in BF16; about 426 MB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 1,885
• Likes: 6
• License: apache-2.0
• Based on: Qwen/Qwen3.5-0.8B
• Quantized versions: 2
• Fine-tunes: 2
• 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/OneJev1.6-0.8B
• Files: https://huggingface.co/OmniJev/OneJev1.6-0.8B/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download OmniJev/OneJev1.6-0.8B

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