OneJev-4B
Multimodal decision model in four sizes (0.8B to 27B), fine-tuned from Qwen3.5 and Qwen3.8. Reads a state of screenshots, photos, video frames or text and returns a probability per option for choice, score and noul questions, from option-letter logits in one pass.
Full fine-tunes for one epoch with the vision tower frozen, on 99,193 questions from GUI-agent runs, images, short and long videos, text and rules, with a cross-entropy plus Brier loss. The state is prefilled once and the cache forked per question, so ten questions on one screenshot take 104 ms on an H200. The qev server speaks TypeSafe's System One API plus a media field; video needs the PyTorch backend. Siblings: OneJev-0.8B (70.0%), OneJev-9B (76.4%), OneJev-27B on Qwen3.8-27B (77.4%) and OneJev-27B-FP8, on the maker's held-out test set. On DecisionBench (a third-party suite, the maker's own run) the 4B scores 68.9 medium and 45.2 hard, below its untrained base on hard, and the maker's own table shows Jev 1.13 ahead on DecisionBench. 94,707 of the training questions are released; the maker's dataset card says some sources are research-only or non-commercial. PlayJev-0.8B, a separate game-playing policy for 10 browser games, is linked below. Third-party GGUF builds exist and are not listed. By Li Bobo.
What it decides
- choice — picks one option from a set
- score — places the input on an ordered scale
- noul — answers a yes/no question with one probability
At a glance
| Parameters | 4B |
| Base model | Qwen/Qwen3.5-4B |
| Maker | OmniJev |
| Released | 2026-09-27 |
| License | apache-2.0 |
| Reported accuracy | 75.7% |
| Reported latency | 64 ms for one question, 104 ms for ten questions about one 1280x720 screenshot on one H200 |
Get the weights
pip install systemonemodels
systemone pull omnijev/onejev
The files are served from the maker's Hugging Face repository, OmniJev/OneJev-4B, and verified against the checksums recorded here.