Blame for README.md

---
license: apache-2.0
base_model: Qwen/Qwen3.5-9B
library_name: transformers
pipeline_tag: image-text-to-text
tags: [onejev, system-one, decision-model, calibration, multimodal, gui-agent, video]
---

<p align="center">
  <picture>
    <source media="(prefers-color-scheme: dark)" srcset="assets/banner-dark.svg">
    <img alt="OneJev1.6-9B, a Multimodal System One Decision Model" src="assets/banner.svg" width="100%">
  </picture>
</p>

<p align="center">
  <a href="https://huggingface.co/OmniJev"><img alt="Hugging Face" src="https://img.shields.io/badge/Hugging_Face-OmniJev-FFD21E?style=flat-square&logo=huggingface&logoColor=black"></a>
  <a href="https://github.com/OmniJev/OneJev"><img alt="GitHub" src="https://img.shields.io/badge/GitHub-OmniJev%2FOneJev-181717?style=flat-square&logo=github&logoColor=white"></a>
  <a href="https://omnijev.github.io/OneJev/"><img alt="Website" src="https://img.shields.io/badge/Website-OneJev-0A84FF?style=flat-square&logo=googlechrome&logoColor=white"></a>
  <a href="https://github.com/OmniJev/OneJev/blob/main/LICENSE"><img alt="License" src="https://img.shields.io/badge/License-Apache_2.0-D22128?style=flat-square&logo=apache&logoColor=white"></a>
</p>

OneJev1.6-9B is the 9B model of OneJev 1.6, a multimodal System One decision model. Give it a screenshot, a photo, a video or plain text along with a few typed questions, and it returns a calibrated probability for every option in a single forward pass. It is a full fine-tune of [Qwen/Qwen3.5-9B](https://huggingface.co/Qwen/Qwen3.5-9B) on about 440,000 questions drawn from real agent runs, videos, images and the training splits of public decision datasets.

## Decision Index

On the public part of [Decision Index 0.3](https://huggingface.co/spaces/multimodalart/jev-decision-index) (37 benchmarks, 140,178 requests, run and scored with the board's own kit), OneJev1.6-9B scores 65.31, up from 59.03 for OneJev1.5-9B. That is sixth of the 172 models on the board, and the five models above it all have 26B parameters or more. It has the best score in four of the five areas among the five models below.

<p align="center">
  <picture>
    <source media="(prefers-color-scheme: dark)" srcset="assets/di-class-table-dark.svg">
    <img alt="Decision Index public part, table: OneJev1.6-9B 65.31 against Nace.AI Drex v1.5 58.08, vLLM-SR d3-flash (9B) 57.79, Standard One 8B SH (v4) 57.40, Bespoke Nimble 9B v3 57.19, overall and in each of the five areas; best value in each row in magenta, second best in light pink" src="assets/di-class-table.svg" width="100%">
  </picture>
</p>

The four other models are the highest-scoring models of 6B to 12B parameters on the board.

It also scores above 52 of the 57 models of 25B to 36B parameters on the board.

<p align="center">
  <picture>
    <source media="(prefers-color-scheme: dark)" srcset="assets/di-larger-dark.svg">
    <img alt="Decision Index public part: OneJev1.6-9B 65.3 at 9.4B parameters against Torchcast Decision 65.1, Amos 1 Large 64.9, Decider 31B 64.3, linne-oak 1.0 63.9, span-02 26B-A4B 63.6, Matilda Jev v1.5 63.5 at 26B to 33B" src="assets/di-larger.svg" width="100%">
  </picture>
</p>

Other models' scores are their public-part scores on the board on 16 October 2026.

## Results

<p align="center">
  <picture>
    <source media="(prefers-color-scheme: dark)" srcset="assets/results-dark.svg">
    <img alt="Results table: OneJev1.6-9B against Jev 1.13, Jev-Omni 12B and Qwen3.8-27B thinking; best score in each row in magenta, second best in light pink" src="assets/results.svg" width="100%">
  </picture>
</p>

Accuracy in percent. The OneJev test set holds questions of the training kinds that no model saw in training. Jev 1.13's scores are its published ones, and it reads text only.

## Quick start

```bash
uv venv -p 3.12 && source .venv/bin/activate
uv pip install flide torchvision
flide serve --model OmniJev/OneJev1.6-9B
```

In another terminal, run this example with your own `screenshot.png`:

```bash
IMG=$(base64 -w0 screenshot.png)
curl http://localhost:8000/v1/systemone -H "Content-Type: application/json" -d @- <<EOF
{
  "state": {"task": "Pay the open invoice from ACME", "screen": "<image:1>"},
  "media": [{"type": "image", "data": "data:image/png;base64,$IMG"}],
  "questions": {
    "done": {"type": "noul", "instructions": "The invoice has been paid"},
    "next": {"type": "choice", "instructions": "What should the agent do next?",
             "criteria": {"click": "click an element", "type": "type text", "scroll": "scroll", "stop": "stop"}},
    "progress": {"type": "score", "instructions": "How far along is the task?",
                 "criteria": ["not started", "halfway", "almost done", "done"]}
  }
}
EOF
```

`answers.done.noul` is the probability of yes, `answers.next.probabilities` gives one probability per option and `answers.progress.score` is the expected level.

[Flide](https://github.com/OmniJev/Flide) ([paper](https://arxiv.org/abs/2610.42148)) serves TypeSafe's System One API plus a `media` field for images and video. More examples are on [GitHub](https://github.com/OmniJev/OneJev).

## Citation

```bibtex
@misc{onejev2026,
  title        = {{OneJev}: A Multimodal System One Decision Model},
  author       = {{OmniJev Team}},
  year         = {2026},
  howpublished = {\url{https://github.com/OmniJev/OneJev}}
}
```

## License

Apache 2.0