OneJev: an open, multimodal System One decision family from 0.8B to 27B
The OmniJev organisation published OneJev on Hugging Face on 27 September 2026 (around 13:15 BRT): four open decision models (0.8B, 4B, 9B and 27B, plus a 27B FP8 build) under Apache-2.0. Each is a full fine-tune of a Qwen vision-language model with the vision tower frozen, trained for one epoch with cross-entropy plus Brier loss on 99,193 questions. 94,707 of them are released as the OneJev-Data dataset.
What is different. Like Jev, OneJev answers typed questions in one prefill pass and returns probabilities. Unlike Jev, the state can include screenshots, photos or video frames. The qev server speaks System One, so the TypeSafe SDK works against it; GGUF builds run on llama.cpp. Two days later the same maintainer released OpenDecisions, a server that puts OneJev behind /v1/decisions and /v1/systemone with an optional fallback to hosted Jev.
The numbers (UnverifiedClaim). The 27B scores 77.4 on the held-out OneJev test set (Jev-Omni 60.8, Qwen3.8-27B thinking 71.6), 87.9 vs 90.5 for Jev on DecisionBench medium and 73.1 vs 65.3 on DecisionBench hard. The Jev numbers are TypeSafe’s published text-only scores. Latency on an H200 with a 1280×720 screenshot runs from 31 ms (0.8B, one question) to 324 ms (27B, ten questions). OneJev is not on the official Decision Index.
Watch out: pip install qev installs an unrelated package; install from the GitHub repo.
How we list it. Catalog, 5.4 maturity / 7.3 capability / 10 adoption.