OneJev is a multimodal System One decision model. It returns calibrated probabilities for typed questions about screenshots, photos, videos and text in a single forward pass. Available in 0.8B, 4B, 9B and 27B.
Accuracy (%). The OneJev test set is held out from OneJev training. Jev 1.13 uses published text-only scores; Jev-Omni and Qwen3.8-27B thinking were evaluated by us.
Choose one backend, then run the Python example below.
For NVIDIA GPUs. Supports text, images and video.
pip install "qev[torch] @ git+https://github.com/OmniJev/OneJev.git"
qev serve --model OmniJev/OneJev-4B --port 8000For GGUF models. Supports text and images; use PyTorch for video. Install
llama.cpp first (brew install llama.cpp on macOS).
pip install git+https://github.com/OmniJev/OneJev.git
qev serve --gguf mradermacher/OneJev-4B-GGUF:Q8_0 --port 8000The repo ships a Compose stack that detects your hardware and starts one backend.
The llama.cpp profiles download mradermacher/OneJev-4B-GGUF:Q8_0 on first start; set QEV_GGUF in docker/.env
for another size or quant.
cp docker/.env.example docker/.env # set HF_TOKEN for gated/private repos
./docker/start.sh # auto-detect (Windows: .\docker\start.ps1)
./docker/start.sh logs # follow logs
./docker/start.sh down # stop every profileForce a backend with ./docker/start.sh <torch|rocm|vulkan|gguf|gguf-rocm|cpu>
(add --build to rebuild). Every profile serves the API and the playground on :8000.
On Windows only torch (NVIDIA) and cpu are available.
Both backends serve the same API at http://localhost:8000. In another terminal, run this example with your own
screenshot.png:
from qev import Client, Choice, Noul, Score
from qev.media import data_uri
r = Client("http://localhost:8000").system_one(
state={"task": "Pay the open invoice from ACME", "screen": "<image:1>"},
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", "type": "type text", "scroll": "scroll", "stop": "stop"}),
"progress": Score("How far along is the task?", ["not started", "halfway", "almost done", "done"])},
)
r.answers["done"].noul # probability of yes
r.answers["next"].probabilities # one probability per option
r.answers["progress"].score # expected levelThe API is compatible with TypeSafe System One. More examples: video, curl, official TypeSafe SDK.
git clone https://github.com/OmniJev/OneJev.git
cd OneJev
pip install -e ".[train]"
# Choose A or B. A is enabled below.
# A. Demo: 100 examples with images (23 MB)
hf download OmniJev/OneJev-Data sample/sample-100.parquet --repo-type dataset --local-dir data/onejev
python -m train.unpack data/onejev --sample
# B. Full dataset: 94,707 examples (17.7 GB). Uncomment these two lines instead of A.
# hf download OmniJev/OneJev-Data --repo-type dataset --include "data/*.parquet" --local-dir data/onejev
# python -m train.unpack data/onejevEither option creates data/onejev/train.jsonl and extracts images and video frames into data/onejev/media/.
OneJev-Data releases 94,707 of the original 99,193 training questions; the remaining sources do not permit redistribution.
The four models fine-tune Qwen3.5-0.8B, Qwen3.5-4B, Qwen3.5-9B and Qwen3.8-27B for one epoch with the vision tower
frozen. The configs use a learning rate of 5e-6, a 16,384-token limit, and cross-entropy plus Brier loss over answer
probabilities. To train the 4B model on four GPUs:
torchrun --nproc-per-node 4 -m train.sft --config train/configs/onejev_4b_full.yamlConfigs: 0.8B · 4B ·
9B · 27B.
Set --nproc-per-node to your GPU count; the 9B and 27B configs enable FSDP to shard model and optimizer state.
For your own data, change train and media_root in the config.
The final 4B checkpoint is saved to train/runs/onejev_4b/final. Start it with:
qev serve --model train/runs/onejev_4b/final --port 8000@misc{onejev2026,
title = {{OneJev}: A Multimodal System One Decision Model},
author = {{OmniJev Team}},
year = {2026},
howpublished = {\url{https://github.com/OmniJev/OneJev}}
}
