OmniJev/OneJev-27B-FP8

OneJev-27B with its decoder weights in 8-bit floating point, one scale per weight row and per token: 30.4 GB instead of 54.7 GB, so it fits on one 48 GB GPU. On 229 test rows it gives the 16-bit model's answer on 226 (98.7%), accuracy 65.5 for 16-bit and 66.4 for 8-bit. It needs a GPU with FP8 (L40S, H100, H200 and newer).

On one H200 with a 1280x720 screenshot it answers 1 question in 168 ms and 10 questions in one request in 298 ms, against 189 ms and 324 ms for 16-bit.

Quick start

pip install "qev[torch] @ git+https://github.com/OmniJev/OneJev.git"
qev serve --model OmniJev/OneJev-27B-FP8

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": "<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", "stop": "stop"})},
)

The server speaks TypeSafe's System One API plus a media field for images and video. More examples, the latency benchmark and the code are on GitHub.

All sizes

All sizes are in the OneJev collection.

Citation

@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