OneJev 1.6 releases new weights for visual decisions
OmniJev has released OneJev 1.6, with downloadable decision models in four sizes from 0.8 billion to 27 billion parameters. The update gives developers new weights for evaluating questions about text and visual inputs.
The project dates its announcement October 17, 2026. Its release commit was recorded on October 16 at 19.10 UTC. Hugging Face histories place the four weight uploads between 18.52 and 20.37 UTC that evening. Those records establish a new checkpoint release, separate from the original OneJev launch in September.
What the model scores
OneJev accepts supplied questions and returns probabilities across defined answers. The team describes the outputs as calibrated and produced in one forward pass. The model card lists screenshots, photographs, video and text as inputs.
For an illustrative support workflow, a developer could supply a screenshot and ask whether an export finished, which error is visible and whether the case needs review. The application still has to define those choices and decide what to do with the scores.
That puts OneJev in the decision model category covered in ByteForward’s Microsoft Decision 1 report. OneJev’s downloadable visual models provide another deployment option for teams evaluating this approach.
What changed in version 1.6
The current 4B card describes a full Qwen fine tune using roughly 440,000 questions. Its 1.5 predecessor lists about 320,000. Both describe examples drawn from agent activity, visual material and training partitions of public decision datasets.
OmniJev reports a public Decision Index score of 62.01 for the new 4B model, versus 54.89 for 1.5. That is a claim about the public portion of the benchmark, not its complete ranking.
The Decision Index methodology gives public tests 20 percent of its full score, private tests of similar skills 50 percent and private tests in new domains 30 percent. The retrieved leaderboard snapshot predates the 1.6 uploads and contains four OneJev 1.5 entries but no 1.6 results. It therefore cannot independently confirm the new release’s scores.
Using probability scores as decision thresholds requires checking how confidence matches observed outcomes in the intended workload. A team should test recorded cases with known answers, including ambiguous screens and missing evidence. For an export checker, confidently declaring a failed export complete deserves separate attention from sending a successful export for unnecessary review.
Weights and training data have different scope
The 4B download includes a roughly 9.08 GB weights file. Its stored file hash differs from the 1.5 checkpoint, confirming a distinct artifact. Download size alone does not establish the memory needed to load and serve the model.
The serving update changes the PyTorch default to 1.6 and adds model aliases. The GGUF instructions still reference a 1.5 checkpoint. Developers following those examples should verify the model actually loaded before comparing versions.
The linked public training dataset contains 94,707 examples from the original release. It excludes material whose sources do not permit redistribution and applies separate licenses by source. That archive does not establish access to the full 440,000 question collection described for 1.6, limiting what an outside team can reproduce from the public training materials alone.
The release cards and code specify Apache 2.0. The underlying Qwen model cards also identify Apache 2.0. These declarations cover the respective model and code artifacts. They do not replace the dataset’s source specific conditions.
The concrete advance is a fresh set of weights with an updated serving path. The next useful evidence is performance on a team’s own screens and questions, with the exact checkpoint recorded and mistakes reviewed before scores trigger actions.
Archival server room photograph at SpinVFX in 2018, used for computing context. Photo by KSingh1991 via Wikimedia Commons under Creative Commons Attribution ShareAlike 4.0. Cropped and resized. The adapted photograph uses the same license.
