Independent JEV-35B brings one-pass decisions to open weights
AutoTrust's Qwen-based model exposes yes/no, score and multiple-choice decisions, but its benchmark and speed claims come from the project itself.
An independent decision model developers can run
AutoTrust AI has published JEV-35B, an Apache-2.0 open-weight model built on Qwen3.5-35B-A3B for typed decisions rather than open-ended chat. The English Hugging Face repository contains checkpoint files, configuration and serving code. Its card describes a single-pass interface for yes/no answers, scores and choices among as many as 256 options. AutoTrust explicitly says the model is independent of TypeSafe AI’s hosted Jev product and is neither affiliated with nor endorsed by that company.
What the evidence shows
Hugging Face records the repository’s creation on October 11 and a source update on October 13. It appeared in an October 13 trend observation, which shows current visibility, not a fresh launch that day or a measured growth rate. AutoTrust reports its own Decision Index result and latency measurements, but labels the public-suite score as its own run rather than an official board entry. The card also says JEV-35B cannot chat or generate prose and performed worse than its comparison model on several game tasks. Developers need their own tests before treating its calibrated outputs as reliable decisions.