OmniJev/OneJev

What it claims

OneJev 1.6, released 17 October 2026 on Hugging Face and ModelScope, is the current version of the model. It follows 1.5 (13 October) and the initial release with the dataset (27 September). It answers several typed questions about one state in a single request. The three types are a yes/no probability, a probability per option for a choice, and an expected level on an ordinal scale for a score. The README's worked example is a GUI agent loop: is the task done, what action comes next, and how far along the task is. The page names an accuracy table on its own held-out OneJev test set plus public benchmarks, but the fetched text does not include the numbers.

How to use it

There are two backends. Flide is the team's own inference engine for NVIDIA GPUs and handles text, images and video:

uv venv -p 3.12 && source .venv/bin/activate
uv pip install flide torchvision
flide serve --model OmniJev/OneJev1.6-0.8B --port 8000

The llama.cpp path uses GGUF quants published by mradermacher. It supports text and images only, not video:

pip install git+https://github.com/OmniJev/OneJev.git
qev serve --gguf mradermacher/OneJev1.6-0.8B-GGUF:Q8_0 --port 8000

The Docker Compose stack auto-detects hardware via ./docker/start.sh, or a profile can be forced with torch|rocm|vulkan|gguf|gguf-rocm|cpu. GGUF profiles pull the 0.8B Q8_0 by default. Set QEV_GGUF in docker/.env to choose another size or quant. Every profile serves the API and a playground on port 8000.

Requests go to POST /v1/systemone. The body has a state object that references media by placeholder ("<image:1>"), a media array of base64 data URIs, and a questions map. Each question has a type (noul, choice or score), instructions, and for choice and score, criteria. Responses come back as answers.<name>.noul (P(yes)), answers.<name>.probabilities and answers.<name>.score. The page states the API is compatible with TypeSafe System One and its official SDK.

Training

The model is fine-tuned for one epoch with the vision tower frozen. Training uses learning rate 5e-6 and a 16,384-token limit. The 9B and 27B configs enable FSDP. The 4B run on four GPUs:

pip install -e ".[train]"
torchrun --nproc-per-node 4 -m train.sft --config train/configs/onejev_4b_full.yaml

A 100-example sample (23 MB) is available for smoke tests. The full set is 17.7 GB. python -m train.unpack writes train.jsonl and extracts images and video frames to media/. Custom data needs only train and media_root changed in the config. Trained checkpoints are served with flide serve --model <path> --format onejev.

Limits

The public OneJev-Data has 94,707 of the original 99,193 training questions. The rest come from sources that do not permit redistribution, so a retrain from public data does not exactly reproduce the released models. Video needs Flide and therefore an NVIDIA GPU. On Windows the Docker stack offers only the torch (NVIDIA) and cpu profiles. Accuracy results include a test set the authors built themselves, and the fetched page does not show the scores. The page states no licence for the weights or the dataset, so both need checking before adoption.