---
title: OneJev1.6-4B-GGUF
canonical_url: "https://www.modelscope.ai/models/OmniJev/OneJev1.6-4B-GGUF"
md_url: "https://www.modelscope.ai/models/OmniJev/OneJev1.6-4B-GGUF.md"
repository: OmniJev/OneJev1.6-4B-GGUF
last_updated: 2026-10-18
license: apache-2.0
pipeline_tag: image-text-to-text
tasks:
- image-text-to-text
base_model:
- OmniJev/OneJev1.6-4B
base_model_relation: quantized
library_name:
- gguf
downloads: 38
stars: 0
tags:
- onejev
- system-one
- decision-model
- calibration
- multimodal
- gui-agent
- gguf
- llama.cpp
---
# OneJev1.6-4B-GGUF
> OneJev1.6-4B-GGUF - An open-source model by OmniJev on ModelScope. This repository holds GGUF files of OneJev1.6-4B for llama.cpp.
OmniJev/OneJev1.6-4B-GGUF is a image-text-to-text model on ModelScope. licensed under apache-2.0. derived from OmniJev/OneJev1.6-4B.
- **Repository**: OmniJev/OneJev1.6-4B-GGUF
- **License**: apache-2.0
- **Tasks**: image-text-to-text
- **Base model**: OmniJev/OneJev1.6-4B
- **Tags**: onejev, system-one, decision-model, calibration, multimodal, gui-agent, gguf, llama.cpp
- **Downloads**: 38
- **Stars**: 0
- **Last updated**: 2026-10-18
Source: https://www.modelscope.ai/models/OmniJev/OneJev1.6-4B-GGUF
---
This repository holds GGUF files of [OneJev1.6-4B](https://modelscope.cn/models/OmniJev/OneJev1.6-4B) for llama.cpp.
OneJev1.6-4B is the 4B 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/Qwen3.5-4B](https://huggingface.co/Qwen/Qwen3.5-4B) on about 440,000 questions drawn from real agent runs, videos, images and the training splits of public decision datasets.
## Files
Converted with [llama.cpp](https://github.com/ggml-org/llama.cpp/commit/23f7aabc16c72f6f707dba227251c35d7bd1ed96) from the released weights of [OneJev1.6-4B](https://modelscope.cn/models/OmniJev/OneJev1.6-4B).
| File | Size | |
|---|---|---|
| `OneJev1.6-4B-Q4_K_M.gguf` | 2.7 GB | smallest, for low memory |
| `OneJev1.6-4B-Q5_K_M.gguf` | 3.1 GB | |
| `OneJev1.6-4B-Q6_K.gguf` | 3.5 GB | |
| `OneJev1.6-4B-Q8_0.gguf` | 4.5 GB | recommended, closest to the original |
| `OneJev1.6-4B-F16.gguf` | 8.4 GB | full precision |
| `mmproj-OneJev1.6-4B-F16.gguf` | 672 MB | vision projector, needed for images |
## 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), OneJev1.6-4B scores 62.01, up from 54.89 for OneJev1.5-4B. Every model on the board that scores higher has 25B parameters or more. It has the best score in four of the five areas among the five models below, and the second best in Language Understanding.
The four other models are the highest-scoring models of 3B to 6B parameters on the board.
It also scores above 42 of the 57 models of 25B to 36B parameters on the board.
Other models' scores are their public-part scores on the board on 16 October 2026.
## Results
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
Install [llama.cpp](https://github.com/ggml-org/llama.cpp/blob/master/docs/install.md) first (`brew install llama.cpp` on macOS), then download one model file and the vision projector and serve them with qev.
```bash
pip install modelscope git+https://github.com/OmniJev/OneJev.git
modelscope download --model OmniJev/OneJev1.6-4B-GGUF OneJev1.6-4B-Q8_0.gguf mmproj-OneJev1.6-4B-F16.gguf --local_dir OneJev1.6-4B-GGUF
qev serve --gguf OneJev1.6-4B-GGUF/OneJev1.6-4B-Q8_0.gguf --port 8000
```
qev finds the vision projector next to the model file and fetches the tokenizer from Hugging Face. Without Hugging Face access, download the tokenizer from ModelScope and pass it with `--tokenizer`.
```bash
modelscope download --model OmniJev/OneJev1.6-4B --include "*.json" "*.jinja" --local_dir OneJev1.6-4B
qev serve --gguf OneJev1.6-4B-GGUF/OneJev1.6-4B-Q8_0.gguf --tokenizer OneJev1.6-4B --port 8000
```
```python
from qev import Client, Choice, Noul
from qev.media import data_uri
r = Client("http://localhost:8000").system_one(
state={"task": "Pay the open invoice from ACME", "screen": ""},
media=[{"type": "image", "data": data_uri("screenshot.png")}],
questions={"done": Noul("The invoice has been paid"),
"next": Choice("What should the agent do next?", {"click": "click an element", "stop": "stop"})},
)
```
The server speaks TypeSafe's System One API plus a `media` field for images. For video, use the PyTorch weights in [OneJev1.6-4B](https://modelscope.cn/models/OmniJev/OneJev1.6-4B). More examples and the code 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
